This book brings together several advanced topics in computer graphics that are important in the areas of game developme

*414*
*173*
*8MB*

*English*
*Pages xiii, 312 pages: illustrations
[316]*
*Year 2012*

- Author / Uploaded
- Mukundan
- Ramakrishnan

*Table of contents : Advanced Methods in Computer Graphics......Page 3Preface......Page 6Contents......Page 8 1.1 Advanced Computer Graphics......Page 13 1.3 Notations......Page 14 1.4 Contents Overview......Page 15 2.1 Points and Vectors......Page 17 2.2 Signed Angle and Area......Page 21 2.3 Lines and Planes......Page 23 2.4 Intersection of 3 Planes......Page 26 2.5 Curves......Page 28 2.6 Affine Transformations......Page 29 2.7 Affine Combinations......Page 31 2.8 Barycentric Coordinates......Page 34 2.9 Basic Lighting......Page 36 2.11 Supplementary Material for Chap. 2......Page 38 References......Page 41 3.1 The Basic Structure of a Scene Graph......Page 43 3.2 Transformation Hierarchy......Page 45 3.2.1 A Mechanical Part......Page 46 3.2.2 A Simple Character Model......Page 47 3.2.3 A Planetary System......Page 48 3.3 Relative Transformations......Page 50 3.4 Bounding Volume Hierarchy......Page 52 3.5.1 Group Node......Page 55 3.5.2 Object Node......Page 56 3.5.4 Light Node......Page 57 3.6 First-Person View......Page 59 3.8 Supplementary Material for Chap. 3......Page 61 3.9 Bibliographical Notes......Page 63 References......Page 64 4.1 Articulated Character Models......Page 65 4.2 Vertex Blending......Page 67 4.3 Skeleton and Skin......Page 69 4.4.1 The Bind Pose......Page 71 4.4.2 Mesh Vertex Transformation......Page 72 4.5 Vertex Skinning Using Scene Graphs......Page 74 4.6 Transformation Blending......Page 76 4.7 Keyframe Animation......Page 78 4.8.2 Skinned Mesh Node......Page 81 4.9 Summary......Page 84 4.10 Supplementary Material for Chap. 4......Page 86 References......Page 88 5.1 Review of Complex Numbers......Page 89 5.2 Quaternion Algebra......Page 91 5.3 Quaternion Transformation......Page 93 5.4 Generalized Rotations......Page 95 5.4.1 Euler Angles......Page 96 5.4.2 Angle-Axis Transformation......Page 98 5.5 Quaternion Rotations......Page 100 5.5.1 Quaternion Transformation Matrix......Page 102 5.5.2 Quaternions and Euler Angles......Page 103 5.5.3 Negative Quaternion......Page 104 5.6 Rotation Interpolation......Page 105 5.6.1 Euler Angle Interpolation......Page 107 5.6.2 Axis-Angle Interpolation......Page 108 5.6.3 Quaternion Linear Interpolation (LERP)......Page 110 5.6.4 Quaternion Spherical Linear Interpolation (SLERP)......Page 111 5.7 Quaternion Exponentiation......Page 113 5.8 Relative Quaternions......Page 114 5.9.1 Dual Numbers......Page 116 5.9.2 Algebra of Dual Quaternions......Page 117 5.9.3 Transformations Using Dual Quaternions......Page 120 5.10 Summary......Page 121 5.11 Supplementary Material for Chap. 5......Page 122 5.12 Bibliographical Notes......Page 123 References......Page 124 6.1 Robot Manipulators......Page 125 6.2.1 Joint Chain in Two Dimensions......Page 127 6.2.2 Joint Chain in 3D Space......Page 128 6.3 Linear and Angular Velocity......Page 130 6.3.1 Velocity in Two Dimensions......Page 131 6.3.2 Velocity Under Euler Angle Transformations......Page 132 6.3.3 Quaternion Velocity......Page 133 6.3.4 The Jacobian......Page 135 6.4 Inverse Kinematics......Page 136 6.4.1 2-Link Inverse Kinematics......Page 137 6.4.2 n-Link Inverse Kinematics......Page 138 6.5 Gradient Descent......Page 140 6.6 Cyclic Coordinate Descent......Page 142 6.7 Circular Alignment Algorithm......Page 144 6.9 Supplementary Material for Chap. 6......Page 147 References......Page 148 7.1 Polynomial Interpolation......Page 150 7.2 Cubic Parametric Curves......Page 152 7.3 Parametric Continuity......Page 156 7.4 Hermite Splines......Page 158 7.5 Cardinal Splines......Page 160 7.6.1 Cubic Bezier Splines......Page 162 7.6.2 de-Casteljau's Algorithm......Page 165 7.7 Polynomial Interpolants......Page 167 7.8 B-Splines......Page 170 7.8.1 Basis Functions......Page 171 7.8.2 Approximating Curves......Page 174 7.8.3 NURBS......Page 177 7.9 Surface Patches......Page 178 7.10 Coons Patches......Page 181 7.11 Bi-Cubic Bezier Patches......Page 183 7.13 Supplementary Material for Chap. 7......Page 185 References......Page 188 8.1 Mesh Representation......Page 189 8.2 Polygonal Manifolds......Page 193 8.3 Mesh Data Structures......Page 196 8.3.1 Face-Based Data Structure......Page 197 8.3.2 Winged-Edge Data Structure......Page 198 8.3.3 Half-Edge Data Structure......Page 200 8.4.1 Vertex Decimation......Page 204 8.4.2 Edge Collapse Operation......Page 206 8.5.1 Subdivision Curves......Page 211 8.5.2 The Loop Subdivision Algorithm......Page 213 8.5.3 Catmull-Clark Subdivision......Page 215 8.5.4 Root-3 Subdivision......Page 217 8.6 Mesh Parameterization......Page 219 8.6.1 Barycentric Embedding......Page 220 8.6.2 Spherical Embedding......Page 224 8.7 Polygon Triangulation......Page 225 8.7.1 Polygon Types......Page 226 8.7.2 Edge-Flip Algorithm......Page 228 8.7.3 Three Coins Algorithm......Page 229 8.7.4 Triangulation of Monotone Polygons......Page 232 8.8 Summary......Page 234 8.9 Supplementary Material for Chap. 8......Page 236 8.10 Bibliographical Notes......Page 238 References......Page 239 9.1 Bounding Volumes......Page 240 9.1.2 Minimal Bounding Sphere......Page 241 9.1.3 Oriented Bounding Box (OBB)......Page 246 9.1.4 Discrete Oriented Polytope (k-DOP)......Page 248 9.1.5 Convex Hulls......Page 250 9.2.1 AABB Intersection......Page 252 9.2.2 OBB Intersection......Page 255 9.2.3 Sphere Intersection......Page 260 9.2.4 k-DOP Intersection......Page 261 9.2.5 Triangle Intersection......Page 262 9.3 Bounding Volume Hierarchies......Page 266 9.3.1 Top-Down Design......Page 267 9.3.2 Bottom-Up Design......Page 268 9.3.3 Collision Testing Using Hierarchy Traversal......Page 269 9.3.4 Cost Function......Page 271 9.4.1 Octrees......Page 272 9.4.2 k-d Trees......Page 276 9.4.3 Boundary Interval Hierarchy......Page 280 9.6 Supplementary Material for Chap. 9......Page 282 9.7 Bibliographical Notes......Page 284 References......Page 285Appendices......Page 286 A.1 Point3 Class......Page 287 A.2 Vec3 Class......Page 289 A.3 Triangle Class......Page 290 A.4 Matrix Class......Page 292 B.1 GroupNode Class......Page 294 B.2 ObjectNode Class......Page 296 B.3 CameraNode Class......Page 297 B.4 LightNode Class......Page 298 C.1 SkeletonNode Class......Page 299 C.2 Skeleton Class......Page 302 C.3 SkinnedMesh Class......Page 303 D.1 Quaternion Class......Page 306 D.2 Dual Quaternion Class......Page 309Index......Page 311*

Advanced Methods in Computer Graphics

Ramakrishnan Mukundan

Advanced Methods in Computer Graphics With examples in OpenGL

123

R. Mukundan Department of Computer Science and Software Engineering University of Canterbury Christchurch, New Zealand

ISBN 978-1-4471-2339-2 e-ISBN 978-1-4471-2340-8 DOI 10.1007/978-1-4471-2340-8 Springer London Dordrecht Heidelberg New York British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Control Number: 2012931936 © Springer-Verlag London Limited 2012 Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms of licenses issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be sent to the publishers. The use of registered names, trademarks, etc., in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant laws and regulations and therefore free for general use. The publisher makes no representation, express or implied, with regard to the accuracy of the information contained in this book and cannot accept any legal responsibility or liability for any errors or omissions that may be made. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

To my daughter Lalitha

Preface

The field of Computer Graphics has evolved rapidly over the past decade following the development of a large collection of algorithms and techniques for various applications in modelling, animation, visualisation, real-time rendering and game engine design. Advances in graphics hardware capabilities and processor technology have continuously fuelled this growth. As a result, this field continues to have enormous potential for further research and development. Computer graphics has also been one of the popular subjects in the computer science and computer engineering disciplines for several years. It is a field where one could always find new and interesting ideas, elegant algorithms and robust implementations. I have been teaching both introductory and advanced courses on computer graphics for the past 12 years, and have constantly observed the enthusiasm of students in learning as well as mastering various techniques used for threedimensional modelling, rendering and animation. The visual effects some of these methods produce captivate their interest, and motivate them to further study and research more advanced techniques. This book evolved from a compilation of my lecture notes and reference material for a graduate course in advanced computer graphics taught in the Department of Computer Science and Software Engineering at the University of Canterbury. The primary aim of this book project has been to develop a reference text suitable for both students and researchers, providing an in-depth and comprehensive coverage of important methods that are useful in the field of character animation. Working towards this goal, I soon realised that a book covering a large number of subfields ranging from physically based simulation to non-photorealistic rendering would be a highly ambitious project. This book includes a selection of topics which I consider as fundamental to the area of animation and rendering, and I hope that it will contribute to a deeper and broader understanding of key algorithms used in advanced computer graphics. I am very much indebted to the graduate students and staff in the Department of Computer Science and Software Engineering, University of Canterbury, for their support, valuable feedback, and encouragement. My sincere thanks go to Dr. Richard Lobb (Adjunct Senior Fellow, Department of Computer Science and Software Engineering, University of Canterbury) for devoting so much of his vii

viii

Preface

valuable time and expertise for reviewing the manuscript. I am thankful to Dr. Christian Long (Department of English, University of Canterbury), for copy-editing the manuscript. His thorough and meticulous checking of spelling, punctuation and grammar has helped improve the clarity of the material presented. I would like to thank the editorial team members for their help throughout this book project. While the manuscript was being prepared, a series of unfortunate events, including the passing away of my mother, and two major earth quakes in Christchurch, brought the progress to a standstill for several months. Special thanks to Helen Desmond and Beverley Ford for their continuous encouragement. They showed a tremendous amount of patience, and always so kindly agreed to extend the manuscript submission deadline a number of times. I am very grateful to my family for their endless support. I greatly appreciate their patience and understanding throughout the time when I was obsessed with writing this book. Department of Computer Science and Software Engineering University of Canterbury Christchurch, New Zealand

R. Mukundan

Contents

1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.1 Advanced Computer Graphics . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.2 Supplementary Material .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.3 Notations .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1.4 Contents Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

1 1 2 2 3

2 Mathematical Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.1 Points and Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.2 Signed Angle and Area . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.3 Lines and Planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.4 Intersection of 3 Planes .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.5 Curves .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.6 Affine Transformations .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.7 Affine Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.8 Barycentric Coordinates .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.9 Basic Lighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.10 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.11 Supplementary Material for Chap. 2 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 2.12 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

5 5 9 11 14 16 17 19 22 24 26 26 29 29

3 Scene Graphs .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.1 The Basic Structure of a Scene Graph .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2 Transformation Hierarchy .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.1 A Mechanical Part . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.2 A Simple Character Model.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.2.3 A Planetary System . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.3 Relative Transformations .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.4 Bounding Volume Hierarchy .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.5 Sample Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.5.1 Group Node .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.5.2 Object Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

31 31 33 34 35 36 38 40 43 43 44 ix

x

Contents

3.5.3 Camera Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.5.4 Light Node .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.6 First-Person View . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.7 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.8 Supplementary Material for Chap. 3 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3.9 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

45 45 47 49 49 51 52

4 Skeletal Animation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.1 Articulated Character Models . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.2 Vertex Blending .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.3 Skeleton and Skin .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4 Vertex Skinning .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4.1 The Bind Pose . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.4.2 Mesh Vertex Transformation.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.5 Vertex Skinning Using Scene Graphs.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.6 Transformation Blending .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.7 Keyframe Animation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.8 Sample Implementation of Vertex Skinning .. . . . .. . . . . . . . . . . . . . . . . . . . 4.8.1 Skeleton Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.8.2 Skinned Mesh Node .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.9 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.10 Supplementary Material for Chap. 4 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4.11 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

53 53 55 57 59 59 60 62 64 66 69 69 69 72 74 76 76

5 Quaternions.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.1 Review of Complex Numbers .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.2 Quaternion Algebra .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.3 Quaternion Transformation . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.4 Generalized Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.4.1 Euler Angles .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.4.2 Angle-Axis Transformation .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.5 Quaternion Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.5.1 Quaternion Transformation Matrix . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.5.2 Quaternions and Euler Angles . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.5.3 Negative Quaternion . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.6 Rotation Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.6.1 Euler Angle Interpolation . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.6.2 Axis-Angle Interpolation.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.6.3 Quaternion Linear Interpolation (LERP) .. . . . . . . . . . . . . . . . . . . . 5.6.4 Quaternion Spherical Linear Interpolation (SLERP) . . . . . . . . 5.7 Quaternion Exponentiation .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.8 Relative Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.9 Dual Quaternions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.9.1 Dual Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

77 77 79 81 83 84 86 88 90 91 92 93 95 96 98 99 101 102 104 104

Contents

xi

5.9.2 Algebra of Dual Quaternions . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.9.3 Transformations Using Dual Quaternions.. . . . . . . . . . . . . . . . . . . 5.10 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.11 Supplementary Material for Chap. 5 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5.12 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

105 108 109 110 111 112

6 Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.1 Robot Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2 Forward Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.1 Joint Chain in Two Dimensions . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.2.2 Joint Chain in 3D Space.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3 Linear and Angular Velocity . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.1 Velocity in Two Dimensions .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.2 Velocity Under Euler Angle Transformations.. . . . . . . . . . . . . . . 6.3.3 Quaternion Velocity . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.3.4 The Jacobian .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.4 Inverse Kinematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.4.1 2-Link Inverse Kinematics .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.4.2 n-Link Inverse Kinematics .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.5 Gradient Descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.6 Cyclic Coordinate Descent .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.7 Circular Alignment Algorithm .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.8 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.9 Supplementary Material for Chap. 6 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6.10 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

113 113 115 115 116 118 119 120 121 123 124 125 126 128 130 132 135 135 136 136

7 Curves and Surfaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.1 Polynomial Interpolation .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.2 Cubic Parametric Curves . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.3 Parametric Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.4 Hermite Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.5 Cardinal Splines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.6 Bezier Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.6.1 Cubic Bezier Splines . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.6.2 de-Casteljau’s Algorithm . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.6.3 Rational Bezier Curves .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.7 Polynomial Interpolants . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.8 B-Splines .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.8.1 Basis Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.8.2 Approximating Curves . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.8.3 NURBS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.9 Surface Patches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.10 Coons Patches.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.11 Bi-Cubic Bezier Patches. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

139 139 141 145 147 149 151 151 154 156 156 159 160 163 166 167 170 172

xii

Contents

7.12 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.13 Supplementary Material for Chap. 7 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7.14 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

174 174 177 177

8 Mesh Processing .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.1 Mesh Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.2 Polygonal Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.3 Mesh Data Structures .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.3.1 Face-Based Data Structure .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.3.2 Winged-Edge Data Structure . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.3.3 Half-Edge Data Structure . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.4 Mesh Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.4.1 Vertex Decimation .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.4.2 Edge Collapse Operation.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.5 Mesh Subdivision .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.5.1 Subdivision Curves .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.5.2 The Loop Subdivision Algorithm . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.5.3 Catmull-Clark Subdivision.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.5.4 Root-3 Subdivision .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.6 Mesh Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.6.1 Barycentric Embedding . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.6.2 Spherical Embedding.. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.7 Polygon Triangulation .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.7.1 Polygon Types .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.7.2 Edge-Flip Algorithm . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.7.3 Three Coins Algorithm.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.7.4 Triangulation of Monotone Polygons . . . .. . . . . . . . . . . . . . . . . . . . 8.8 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.9 Supplementary Material for Chap. 8 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8.10 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

179 179 183 186 187 188 190 194 194 196 201 201 203 205 207 209 210 214 215 216 218 219 222 224 226 228 229

9 Collision Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.1 Bounding Volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.1.1 Axis Aligned Bounding Box (AABB) . . .. . . . . . . . . . . . . . . . . . . . 9.1.2 Minimal Bounding Sphere .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.1.3 Oriented Bounding Box (OBB). . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.1.4 Discrete Oriented Polytope (k-DOP) . . . . .. . . . . . . . . . . . . . . . . . . . 9.1.5 Convex Hulls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2 Intersection Testing .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2.1 AABB Intersection . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2.2 OBB Intersection . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2.3 Sphere Intersection . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2.4 k-DOP Intersection .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.2.5 Triangle Intersection . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

231 231 232 232 237 239 241 243 243 246 251 252 253

Contents

9.3

Bounding Volume Hierarchies . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.3.1 Top-Down Design . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.3.2 Bottom-Up Design . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.3.3 Collision Testing Using Hierarchy Traversal . . . . . . . . . . . . . . . . 9.3.4 Cost Function .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.4 Spatial Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.4.1 Octrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.4.2 k-d Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.4.3 Boundary Interval Hierarchy.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.5 Summary .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.6 Supplementary Material for Chap. 9 . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9.7 Bibliographical Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

xiii

257 258 259 260 262 263 263 267 271 273 273 275 276

Appendices . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 277 Appendix A: Geometry Classes. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . A.1 Point3 Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . A.2 Vec3 Class. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . A.3 Triangle Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . A.4 Matrix Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

279 279 281 282 284

Appendix B: Scene Graph Classes . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . B.1 GroupNode Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . B.2 ObjectNode Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . B.3 CameraNode Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . B.4 LightNode Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

287 287 289 290 291

Appendix C: Vertex Skinning Classes . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . C.1 SkeletonNode Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . C.2 Skeleton Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . C.3 SkinnedMesh Class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .

293 293 296 297

Appendix D: Quaternion Classes . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301 D.1 Quaternion Class. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301 D.2 Dual Quaternion Class. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 304 Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 307

Chapter 1

Introduction

1.1 Advanced Computer Graphics Computer graphics algorithms are being increasingly used in many scientific and technological areas, with an explosive growth in applications requiring threedimensional rendering and animation. The expansion of computer graphics into diverse and interdisciplinary areas is the result of many factors such as the ever increasing power and capability of the graphics hardware, decreasing hardware costs, availability of a wide range of software tools, research advancements in the field, and significant improvements in graphics application programming interface (API). Additionally, vast amounts of resources including images, 3D models, and libraries are now easily available to developers and researchers for their work. With the emergence of programmable graphics hardware, the power of graphics APIs to render complex models and scenes has greatly increased, and it has become easier to create faster and robust implementations of several advanced algorithms. Following these developments, there is also an increasing need for reference books that give an in-depth coverage of advanced methods that are fundamental to many application domains. Advanced computer graphics is a field that encompasses a vast range of topics and a large number of subfields such as game engine development, real-time rendering, global illumination methods and non-photorealistic rendering. Indeed, this field includes a large body of concepts and algorithms not generally covered in introductory graphics texts that deal primarily with basic transformations, projections, lighting, three-dimensional modelling techniques, texturing and rasterization algorithms. This book aims to provide a comprehensive treatment of the theoretical concepts and associated methods related to four core areas: articulated character animation, curve and surface design, mesh processing, and collision detection. The area of character animation is further subdivided into scene graphs, skeletal animation, quaternion rotations and kinematics. A principal objective of this book is to serve as a reference text for both students and researchers. It is designed for courses that build R. Mukundan, Advanced Methods in Computer Graphics: With examples in OpenGL, DOI 10.1007/978-1-4471-2340-8 1, © Springer-Verlag London Limited 2012

1

2

1 Introduction

upon introductory computer graphics concepts. The topics discussed in the book are commonly covered in graduate or advanced undergraduate graphics courses. These include the theoretical as well as the implementation aspects of several algorithms. To help students understand the concepts clearly, a set of demonstration programs is included with each chapter. Necessary class libraries giving the implementations of important methods of each class are also provided. Some of the concepts that have recently found a great deal of importance in research such as dual quaternion transformations, and bounding interval hierarchies are also presented.

1.2 Supplementary Material Each chapter is accompanied by a collection of software modules and demonstration programs that show the details and working of key algorithms. All programs are written in CCC. The reader is assumed to be familiar with the basic OpenGL library, which is a easy-to-program, widely accepted cross platform API for developing graphics applications. To keep the implementations simple, shader language functions or any other OpenGL extensions are not used. The source codes including relevant class definitions and input files can be downloaded from Springer’s website, http://extras.springer.com/978-1-4471-2339-2. The programs are written entirely by the author, with the primary aim of motivating students to explore further each technique, and to implement their own creative ideas. They are just tools which developers and researchers could use to build larger frameworks or to try better solutions. A simple programming approach is used so that students with minimal knowledge of C/CCC language and OpenGL will be able to start using the code and work towards more complex or useful applications. None of the software is optimized in terms of algorithm performance or speed. Similarly, object oriented programming concepts are not heavily used, leaving room for a lot of further development.

1.3 Notations In order to have a clear distinction between points, vectors and other mathematical entities, the following notation is normally used in this book. Note that in exceptional cases, a different notation may be used in each of the following categories to avoid ambiguity. For example, a tangent vector to a curve may be denoted by T(t) instead of t(t). Point: A point is generally denoted by an uppercase letter in italics as P. The threedimensional coordinates of P will be written as (xp , yp , zp ). The vector representation of P having the same components as above will be denoted as p. The coordinates of the point P1 will be written as either (xp1 , yp1 , zp1 ) or, if there is no ambiguity, as simply (x1 , y1 , z1 ).

1.4 Contents Overview

3

Vector: A vector will be denoted by a lowercase letter in italics and bold font as v. Its vector components will be noted as (xv , yv , zv ). Complex number: Complex numbers are treated as two-dimensional vectors and denoted using a lowercase letter in italics and bold font as z. Quaternions: Uppercase letters in italic font (such as Q) will be used to denote quaternions. Dual-quaternions will be denoted using uppercase letters in bold and italic font as Q. Line segment: A line segment will be noted using its end points as AB. Triangle: A triangle will be denoted using its vertices as ABC and its area as ABC. A triangle may also be named using an uppercase letter in italics as T. Plane: Uppercase Greek symbols such as , …, will be used for denoting planes and general polygonal surface elements. Matrices: Matrices will be denoted using uppercase letters in bold font as M.

1.4 Contents Overview This section gives an outline of subsequent chapters of the book. Chapter 2 should be treated as revision material on analytical properties of geometrical primitives and may be skipped if you have a good mathematical background. Chapters 3, 4, 5, 6 are closely related to the area of character animation. Chapters 7, 8, 9 deal with mutually independent topics, and can be read separately in any order. Chapter 2 – Mathematical Preliminaries: This chapter outlines important mathematical concepts related to points, vectors, transformations, lines and planes that are fundamental to several methods in computer graphics. Subsequent chapters in the book make use of the results presented here. Chapter 3 – Scene Graphs: This chapter introduces scene graphs and gives examples to show their importance in representing transformation hierarchies in articulated models. A sample implementation of the basic scene graph structure is provided. Chapter 4 – Skeletal Animation: This chapter discusses the animation of two different types of articulated character models. The processes of vertex blending, vertex skinning and keyframing are introduced. The chapter also gives a sample implementation of a skeleton animation module. Chapter 5 – Quaternions: Quaternions are extensively used in animations to represent three-dimensional rotations. This chapter gives a comprehensive coverage of quaternion algebra, transformations and quaternion based methods for rotation interpolation. A recently introduced concept of dual quaternions is also presented.

4

1 Introduction

Chapter 6 – Kinematics: This chapter presents forward and inverse kinematics solutions for animating a joint chain. Iterative algorithms suitable for graphics applications are also presented. Chapter 7 – Curves and Surfaces: This chapter gives an in-depth treatment of parametric curves, splines and polynomial interpolants. Fundamental techniques in curve and surface design using Hermite splines, cardinal splines and B-splines are presented in detail. Chapter 8 – Mesh Processing: This chapter discusses mesh data structures and algorithms. Important edge-based data structures useful for processing adjacency queries are introduced. Algorithms for mesh simplification, subdivision and parameterization are presented. The chapter also outlines methods for polygon triangulation, which is generally a key component of mesh processing algorithms. Chapter 9 – Collision Detection: This chapter details commonly used bounding volume representations of objects in collision detection algorithms, and presents the computation of bounding volume overlap tests. Bounding volume hierarchies and spatial partitioning trees are also discussed in detail.

Chapter 2

Mathematical Preliminaries

Overview Mathematical operations on points, vectors and matrices are needed for processing information related to geometrical objects. Even in the modelling of a simple threedimensional scene, vectors and matrices play an important role in specifying an object’s position, orientation and transformations. Methods for lighting, intersection testing, projections, etc., use a series of vector operations. This chapter gives an overview of computations using geometrical primitives and shapes that form the basis for several algorithms presented in subsequent chapters of the book. Parametric representations are often used in methods involving geometrical primitives. This chapter deals with analytical equations of lines, planes and curves, and their applications in geometrical computations. Properties of three-dimensional transformations are discussed using their matrix representations. The chapter also introduces concepts such as signed area and distance, affine combinations of points and barycentric coordinates.

2.1 Points and Vectors A point is the most fundamental graphics primitive, and is represented in a threedimensional Cartesian coordinate system by the 3-tuple (x, y, z), where x, y, z denote the distances of the point from the origin of the system along the respective axes directions. In graphics, we commonly use an extended coordinate system, where the same point is denoted by the 4-tuple (x, y, z, 1). This representation is called the homogeneous coordinate system. Homogeneous coordinates provide a unified and elegant framework for representing different types of transformations and projections that are commonly applied to both points and vectors (Box 2.1).

R. Mukundan, Advanced Methods in Computer Graphics: With examples in OpenGL, DOI 10.1007/978-1-4471-2340-8 2, © Springer-Verlag London Limited 2012

5

6

2 Mathematical Preliminaries

Box 2.1 Homogeneous Coordinate System A 3D point given by homogeneous coordinates (a, b, c, d) where d is nonzero, has an equivalent representation in Cartesian coordinates given by (a/d, b/d, c/d). The 4-tuple (a, b, c, 0) denotes a point at infinity that has associated with it a directional vector (a, b, c). The many-one mapping from homogeneous to Cartesian space is shown below: (hx, hy, hz, h) ) 3D Point (x, y, z) for all non-zero values of h. (x, y, z, w) ) 3D Point (x/w, y/w, z/w) if w ¤ 0. (x, y, z, 0) ) 3D Vector (x, y, z).

a

y

b

P-Q

y

y

p+q

P

P Q

c

P+Q

Q

p q

z

x

x

z

z

x

Fig. 2.1 Geometric interpretation of (a) subtraction of a point from another, (b) addition of two points given in homogeneous coordinates, and (c) addition of two vectors

We will now look at the geometrical interpretations of operations of addition and subtraction on homogeneous coordinates. When we subtract a point Q D (xq , yq , zq , 1) from the point P D (xp , yp , zp , 1), we get a vector PQ which has components (xp xq , yp yq , zp zq , 0). This vector originates from the point Q and is directed !

towards the point P, and is denoted as QP . The direct addition of two points P and Q is not a geometrically valid operation, as it can produce different results depending on the coordinate reference frame used. If we use the homogeneous coordinate representation of P and Q as given above, the operation of addition yields (xp C xq , yp C yq , zp C zq , 2), which is actually the midpoint of the line segment PQ (Fig. 2.1b). Points can, however, be added in a special way called the affine combination (see Sect. 2.7) that gives a well-defined point. The addition of two vectors p D (xp , yp , zp , 0) and q D (xq , yq , zq , 0) is always a valid operation that produces another vector p C q D (xp C xq , yp C yq , zp C zq , 0). This vector is along the diagonal of the parallelogram formed by p and q.

2.1 Points and Vectors

7

a

b

B

u•v= cosq |u×v|= 2( ABC)

c

∇

(s•n)n

u θ

s A

C

s

n

r

n

v u×v

Fig. 2.2 (a) Dot-product and cross-product of two vectors u,v. (b) Projection of a vector s on a unit vector u. (c) Reflection of a vector s with respect to a unit vector n Fig. 2.3 The normal vector and area of a triangle specified using vertex coordinates can be computed with the help of two vectors defined along the edges

C n

A

v

B u

Like addition, the operations of negation and scalar multiplication should also be carefully performed on points represented in homogeneous coordinates. It can be seen that the operation of negation given by P D (xp , yp , zp , 1) in effect yields the same point P. In general, the operation of scalar multiplication defined as sP D (sxp , syp , szp , s) for any non-zero value of s, gives the same point P. We will often require the computation of angles between two vectors. This and other operations, such as projection, require vectors to be normalized first. The normalization of a vector is the process of converting it to a unit vector that has a magnitude 1. In order to normalize a vector p D (xp , yp , zp , 0), we simply divide each element by the vector magnitude d given by d D jpj D

q

xp2 C yp2 C z2p

(2.1)

If v is a two-dimensional vector (xv , yv ), then the vector v? D (yv , xv ) is perpendicular to and on the left side of v. The vector v? is sometimes called the perp-vector. It may be noted that v?? D (xv , yv ) D v. Two important vector operations used in graphics are the dot-product and the cross-product. Given two unit vectors u D (xu , yu , zu , 0) and v D (xv , yv , zv , 0), their dot-product u•v D xu xv C yu yv C zu zv is equal to the cosine of the angle between the vectors. The cross-product u v D (yu zv yv zu , zu xv zv xu , xu yv xv yu , 0) is a vector perpendicular to both u and v, so that u, v, u v form a right-handed system (Fig. 2.2). Obviously, this operation is useful for computing the surface normal vector of a planar element defined by two vectors u and v. The magnitude of u v (denoted by ju vj) gives twice the area of the triangle formed by the two vectors (Figs. 2.2a and 2.3). For unit vectors, ju vj is also equal to the sine of the angle between the two vectors (Box 2.2).

8

2 Mathematical Preliminaries

Box 2.2 Vector Products The following facts are commonly used in computations involving vectors: If u is a unit vector, then u•u D 1. If u is perpendicular to v, then u•v D 0. If u is parallel to v, then u v D 0. In particular, u u D 0. The magnitude of u v is the area of the parallelogram formed by u, v. The scalar triple product u•(v w) gives the volume of the parallelepiped formed by the vectors u,v and w. The value does not change with a cyclic permutation of the vectors: u•(v w) D v•(wˇ u) D w•(u ˇ v). ˇ xu yu zu ˇ ˇ ˇ u•(v w) can be written as the determinant ˇˇ xv yv zv ˇˇ ˇx y z ˇ w w w The vector triple product u (v w) is the same as (u•w)v (u•v)w. The magnitudes of the dot and cross products of two vectors u and v are related by the equation: ju vj2 D juj2 jvj2 (u•v)2 .

We saw in the previous paragraph that both the dot and the cross products of two unit vectors can give us the information about the angle between them in the form of trigonometric functions cos() and sin() respectively. Note that the function acos(u•v) returns the angle in the range [0, ] only. Neither can we use asin(ju vj) to determine the angle correctly because the resulting value will always be in the restricted range [0, /2] (even though asin() returns a value in the range [ /2, /2], since ju vj is always positive, so would be the result). We will explore ways to compute the true angle in the range [ , ] in Sect. 2.2. If we represent the vertices of a triangle by points A D (xa , ya , za ), B D (xb , yb , zb ), C D (xc , yc , zc ), the surface normal vector and the area of the triangle can be obtained from the cross product of two vectors u, v constructed as shown in Fig. 2.3. The normal vector n of the triangle in Fig. 2.3 has components (xn , yn , zn ) given by xn D ya .zb zc / C yb .zc za / C yc .za zb / yn D za .xb xc / C zb .xc xa / C zc .xa xb / zn D xa .yb yc / C xb .yc ya / C xc .ya yb /

(2.2)

The above vector is the same as u v. The area of the triangle ABC can be computed from the above components of the normal vector as follows: ABC D

1 2

q

xn2 C yn2 C z2n D

1 ju vj 2

(2.3)

2.2 Signed Angle and Area

9

Let us turn our attention to another important vector operation called projection. A vector s can be projected onto a unit vector n, with the projected vector given by (s•n)n (see Fig. 2.2b). This also implies that the length of the projection of s on a unit vector n is s•n. We can use this fact to express any vector s in terms of its projections along three mutually orthogonal unit vectors u,v, and w as s D .s u/u C .s v/v C .s w/w

(2.4)

If s is also a unit vector, then the terms s•u, s•v, s•w are called the direction cosines of the vector in the coordinate space spanned by the unit vectors u, v, and w. In a new coordinate space defined by u, v, and w, the components of any vector s are therefore given by (s•u, s•v, s•w). The reflection of the vector s with respect to a unit vector n is the vector r that lies on the plane containing s and n as shown in Fig. 2.2c, such that the angle between r and n is the same as the angle between s and n. The reflection vector is commonly used in lighting calculations and ray tracing, where s stands for the vector towards a light source, and n is the surface normal vector. The vector components of r can be computed using the formula r D 2.s n/n s

(2.5)

2.2 Signed Angle and Area In the previous section, we noted that the computation of the angle between two vectors using acos() or asin() functions always yielded only positive values in the range [0, ]. One may suggest using the function atan2(ju vj, u•v). This form of computation of angle has the advantage that neither u nor v needs to be normalized. However, this function also returns values in the positive range [0, ] only, because the numerator ju vj is always positive. The difference between the positive and negative sense of angle is completely view dependent. For vectors residing on the two-dimensional xy-plane, the direction to the viewer is always implied to be the C z direction. In a general three-dimensional case, we need to specify this view direction in order to determine the signed angle in the range [ , ] between two given vectors. If we denote the view direction by w (Fig. 2.4), the angle measured from u to v is positive if the sense of rotation from u to v is anticlockwise when viewed from w. In other words, if w is in the same direction as u v, then the angle is positive, otherwise negative. We can now define the signed angle between u and v with respect to the view vector w as D sign..u v/ w/:cos1

uv jujjvj

(2.6)

10

2 Mathematical Preliminaries C B v

u q

For this view direction, both angle and area are negative.

A u×v w

For this view direction, both angle and area are positive.

Fig. 2.4 The angle between two vectors and the area of the triangle formed by the vectors can have either a positive or a negative sign depending on the orientation of the vertices with respect to a given direction

If u and v are two-dimensional vectors on the xy-plane, we can have the following simplified form for the signed angle: D atan2.xu yv xv yu ; xu yu C xv yv /

(2.7)

We can also define a view-dependent sign for the area of a triangle based on the above concept. If the view vector w has components (xw , yw , zw , 0), Eq. 2.3 now gets modified as follows: q 1 ABC D fsign.xn xw C yn yw C zn zw /g xn2 C yn2 C z2n 2 1 D sign.n w/ (2.8) ju vj 2 where xn , yn , zn are computed from the vertex coordinates using Eq. 2.2. For a triangle on the xy-plane, the right-hand side of the above equation reduces to zn /2. Thus the signed area of a triangle with vertices A D (xa , ya ), B D (xb , yb ), C D (xc , yc ) is ABC D

1 .xa .yb yc / C xb .yc ya / C xc .ya yb // 2

(2.9)

The signed area is positive only if the vertices A, B, C are oriented in an anticlockwise sense with respect to the view direction. The signed area of a triangle is useful in determining if a point is inside the triangle or not. This method is discussed in detail in Sect. 2.8. The concepts presented above are also used for

2.3 Lines and Planes

11

defining the orientation of three points. Three points A, B, C are said to be oriented in the anticlockwise sense with respect a direction w if ..B A/ .C A// w > 0:

(2.10)

If the above condition is satisfied, the three points are said to make a left turn when viewed from the direction w. With reference to Fig. 2.4, the equivalent condition in vector notation is (u v)•w > 0. On the xy-plane, the three points make a left turn if xa .yb yc / C xb .yc ya / C xc .ya yb / > 0:

(2.11)

The reversal of the inequality implies a right turn. The points are collinear if the above expression yields 0. In the next section we will use vector notations and related operations to get concise forms of line and plane equations.

2.3 Lines and Planes Lines and planes form integral parts of three-dimensional models and virtual worlds. A good understanding of line and plane equations and their analytical properties is essential for the development of many applications. For example, even a simple ray tracing application requires the computation of several line-plane intersections. A straight line segment can be defined using two points, say P D (xp , yp , zp , 1) and Q D (xq , yq , zq , 1). The equation of this line in terms of a single parameter t can be expressed as x D xp C t.xq xp /I y D yp C t.yq yp /I z D zp C t.zq zp /

(2.12)

For any value of t between 0 and 1, the above set of equations gives the coordinates of a point on the straight line that lies between P and Q. We can also write the equation of this line segment using vector notation as follows: r D p C tm;

0 t 1:

(2.13)

where r D (x, y, z, 1), p D (xp , yp , zp , 1) and m D QP. The above equation can also be used to represent a ray starting from the point p and having a direction given by the vector m. In this representation, m is generally a unit vector and t can have any positive value. The line given in Eq. 2.12 can be rewritten in the standard form by eliminating t: x xp y yp z zp D D xq xp yq yp zq zp

(2.14)

12

2 Mathematical Preliminaries

Fig. 2.5 Computation of shortest distances of a point V from (a) a line PQ and (b) a plane PQR

b

a V

n V D

D

P S

Q

P

R

Q

From the above equation, we immediately get the condition for the collinearity of three points P D (xp , yp , zp , 1), Q D (xq , yq , zq , 1) and R D (xr , yr , zr , 1): xr xp yr yp zr zp D D xq xp yq yp zq zp

(2.15)

Using Eq. 2.12, we can determine the point S on the line PQ that lies closest to a general three-dimensional point V D (xv , yv , zv , 1). The shortest distance of the point V from the line is given by VS (Fig. 2.5), where S is the projection of the point V on PQ. The point S satisfies the condition that the line segments PQ and VS are orthogonal to each other. Using this condition, the parametric value t of the point S can be obtained as follows: tD

.xv xp /.xq xp / C .yv yp /.yq yp / C .zv zp /.zq zp / .xq xp /2 C .yq yp /2 C .zq zp /2

(2.16)

Substitution of the above value in Eq. 2.12 gives the coordinates of the point S. The shortest (or the perpendicular) distance D of the point V from the line PS is obtained as the distance jVSj. A plane in three-dimensional space is uniquely defined by three non-collinear points, or equivalently, by a point P that lies on the plane and its surface normal vector n. The equation of the plane in terms of the coordinates of the three points P D (xp , yp , zp , 1), Q D (xq , yq , zq , 1), R D (xr , yr , zr , 1), is given by the determinant ˇ ˇx ˇ ˇ xp ˇ ˇ xq ˇ ˇx r

y yp yq yr

z zp zq zr

ˇ 1 ˇˇ 1 ˇˇ D 0: 1 ˇˇ 1ˇ

(2.17)

From this equation of the plane, we get the condition for the coplanarity of four points P, Q, R, S: ˇ ˇ xp ˇ ˇ xq ˇ ˇ xr ˇ ˇx s

yp yq yr ys

zp zq zr zs

ˇ 1 ˇˇ 1 ˇˇ D 0: 1 ˇˇ 1ˇ

(2.18)

2.3 Lines and Planes

13

The determinant is equivalent to (PQ)•(r s) C (RS)•(p q). The condition in Eq. 2.18 also points to the fact that the vectors (QP) and (RS) are coplanar. Thus we can rewrite the above equation using the following scalar triple product: .R P / f.Q P / .S R/g D 0:

(2.19)

The surface normal vector n for the above plane can be obtained (similar to Eq. 2.2), by taking the cross-product of vectors QP and RP. The components of n written as a column vector are given below: 2

3 2 3 xn .yq yp /.zr zp / .yr yp /.zq zp / 6 yn 7 6 .zq zp /.xr xp / .zr zp /.xq xp / 7 6 7D6 7 4 zn 5 4 .xq xp /.yr yp / .xr xp /.yq yp / 5 0

(2.20)

0

The plane equation can be written in point-normal form as .x xp /xn C .y yp /yn C .z zp /zn D 0

(2.21)

which can always be simplified into a linear equation ax C by C cz C d D 0, or expressed using vector notation as .r p/ n D 0;

or equivalently; r n D d;

(2.22)

where d D p•n. The point of intersection of this plane and a ray can be obtained by substituting the equation of the ray, r D q C t m, in the above equation and solving for t. tD

.q n/ d mn

(2.23)

The denominator in the above equation becomes zero when the line is orthogonal to n, i.e., parallel to the plane. The shortest distance D of the point v from the plane (see Fig. 2.5b) is given by the equation DD

.xv xp /xn C .yv yp /yn C .zv zp /zn .v n/ C d p D jnj xn2 C yn2 C z2n

(2.24)

The above term is also called the signed distance of the point v from the plane, as it assumes a positive value if v is on the same side as n, and a negative value otherwise. In general, if the plane’s equation is given in the normal form ax C by C cz C d D 0, where a2 C b2 C c2 D 1, the signed distance of the point v D (xv , yv , zv ) is given by D D axv C byv C czv C d

(2.25)

14

2 Mathematical Preliminaries

R

(s= 0, t= 1)

v (s= 1, t= 0) P u (s= 0, t= 0)

Q r=P +su+ tv

Fig. 2.6 Two-parameter representation of a plane

The above expression can be thought of as the dot product between the vector (a, b, c, d) and (xv , yv , zv , 1), which is the homogeneous representation of v. Note that the unit normal vector to the plane is given by (a, b, c). Signed distances are extensively used in collision detection and point inclusion tests using bounding volumes. Given three non-collinear points P, Q, R, we can have a parametric representation of the plane through the points as r D P C s.Q P / C t.R P / D P C s u C tv

(2.26)

where u and v are vectors along two sides of the triangle PQR (Fig. 2.6). An alternate form for the above equation that expresses any point on the plane as a linear combination of the vertices of the triangle is r D P .1 s t/ C s Q C tR

(2.27)

For every point r(s, t) inside the triangle, the following properties hold: 0 s 1;

0 t 1;

0 s C t 1:

(2.28)

In addition to the above conditions, points along the edge PQ satisfy the parametric equation t D 0. Similarly, the edge PR is characterized by the equation s D 0, and RQ by the property s C t D 1.

2.4 Intersection of 3 Planes An interesting problem commonly encountered while working with planes is the computation of the point of intersection (if it exists) where three planes meet. Even if it is guaranteed that no two planes are parallel, there can be three different configurations in which three planes can meet (Fig. 2.7).

2.4 Intersection of 3 Planes

15

Fig. 2.7 Three different configurations in which three non-parallel planes can meet

In the first configuration in Fig. 2.7, the lines of intersection formed by taking two planes at a time coincide with the result that we get a single line of intersection. In the second configuration, the lines of intersections are parallel even though the planes are not. It can be easily proven that if two lines of intersection are parallel, then the third is also parallel to the other two. This situation arises when the three surface normal vectors of the planes are all coplanar. In the third configuration, the non-parallel lines of intersection meet at a single point. Let the three planes be given by the equations (see Eq. 2.22) r•ni D di , (i D 1, 2, 3) where ni s are unit normal vectors. The directions of the three lines of intersection are then specified by the cross products n1 n2 , n2 n3 , and n3 n1 . The point of intersection, if it exists, can be expressed as a linear combination of these three vectors (Goldman 1990): p D a.n1 n2 / C b.n2 n3 / C c.n3 n1 /

(2.29)

The above point lies on all three planes. Substitution in the plane equations gives bfn1 .n2 n3 /g D d1 cfn2 .n3 n1 /g D d2 afn3 .n1 n2 /g D d3

(2.30)

The scalar triple products on the left side of the above equations are all equal (see Box 2.2). Equation 2.29 can now be written as pD

d1 .n1 n2 / d2 .n2 n3 / d3 .n3 n1 / n1 .n2 n3 /

(2.31)

For the first two configurations shown in Fig. 2.7, the vectors n1 , n2 , n3 are coplanar, and the denominator of the above equation becomes zero. For the third configuration, the equation returns a valid point.

16

2 Mathematical Preliminaries

2.5 Curves In Sect. 2.3, we came across the equation of a straight line expressed in terms of linear polynomials of a single parameter t (Eq. 2.12). Polynomials of a higher degree in t can be used to define curves in three-dimensional space. In the most general form, a curve can be represented as P(t) D (x(t), y(t), z(t)), where x(t), y(t), z(t) are continuous and differentiable functions of the parameter t. Polynomials of degree n have the property that their derivatives up to order n1 exist and are continuous over any finite interval in the parameter space. We can use the derivatives of the functions to define the tangential and normal directions to the curve at any point, and also to construct an orthonormal basis at any point on the curve. The tangent vector at P(t) is given by the first derivative with respect to t, i.e., P0 (t) D (x0 (t), y0 (t), z0 (t)). The unit tangent vector is denoted as T .t/ D

P 0 .t/ jP 0 .t/j

(2.32)

The tangent vector represents the local orientation of the curve at a point. If the parameter t denotes time, then P0 (t) represents the instantaneous velocity of the moving point P(t). The distance travelled from a starting point A D P(t0 ) to the current point, or in other words the arc length measured from A, is given by Zt s.t/ D t0

ˇ 0 ˇ ˇP .u/ˇ d u D

Zt q

.x 0 .u//2 C .y 0 .u//2 C .z0 .u//2 d u

(2.33)

t0

Using the above equation we can express t as a function of arc length s, and re-parameterize the curve as P(s) D (x(s), y(s), z(s)). The chain rule for differentiation gives P 0 .t/ D P 0 .s/s 0 .t/ D P 0 .s/jP 0 .t/j

(2.34)

from which we find that P0 (s) is equivalent to the unit tangent vector T(t). For convenience, we denote P0 (s) by T(s). Since T(s)•T(s) D 1, it immediately follows that T(s)•T0 (s) D 0. Thus the instantaneous rate of change of the tangent direction is parallel to the normal vector at that point. If the unit normal direction at P(s) is denoted as N(s), we have T 0 .s/ D

d .T .s// D .s/N .s/ ds

(2.35)

The proportionality factor (s) is called the curvature of the curve at P(s). The curvature is a measure of the deviation of the curve from a straight line. For a straight line, (s) D 0 at all points. The magnitude of the curvature is easily obtained as j(s)j D jT0 (s)j, and the unit normal direction at P(s) is given by

2.6 Affine Transformations

17

Fig. 2.8 Frenet frame attached to a curve at the point P

P

B(s)

Rectifying plane N(s)

Curve

N .s/ D

T(s)

Osculating plane

P 0 .t/ .P 00 .t/ P 0 .t// P 00 .s/ D 00 jP .s/j jP 0 .t/j jP 00 .t/ P 0 .t/j

(2.36)

The plane containing the tangent vector and the normal vector is known as the osculating plane. The cross-product of the two unit vectors T(s) and N(s) gives the direction of the unit bi-normal vector denoted by B(s): B.s/ D T .s/ N .s/ D

P 0 .t/ P 00 .t/ P 0 .s/ P 00 .s/ D jP 0 .s/ P 00 .s/j jP 0 .t/ P 00 .t/j

(2.37)

The three unit vectors T, N, B form an orthonormal basis as shown in Fig. 2.8. This local reference system is called the Frenet frame. The derivative of the binormal vector B0 (s) is perpendicular to both B(s) and T(s), and hence parallel to N(s): B 0 .s/ D

d .B.s// D .s/N .s/ ds

(2.38)

The term (s) is called the torsion of the curve at s. Torsion is a measure of how much the curve deviates from the osculating plane. The plane containing the tangent and binormal vectors is called the rectifying plane (Fig. 2.8). The plane formed by the normal and binormal vectors is called the normal plane. The Frenet frame is useful for defining the local orientation of objects that move along a curved path. It can also be used for defining the eye-coordinate system for a camera that undergoes a curvilinear motion.

2.6 Affine Transformations In this section, we consider linear transformations of three-dimensional points and vectors. The homogeneous coordinate system (Sect. 2.1) allows all transformations including translations to be represented using 4 4 matrices. We denote a translation

18

2 Mathematical Preliminaries

by a vector v D (xv , yv , zv ), by Tv , a rotation about the x-axis by an angle by ™, by R™ (x), and a scaling by a vector k D (xk , yk , zk ), by Sk (Box 2.3).

Box 2.3 Fundamental 3D Transformations (Fig. 2.9) 3 2 1 0 0 xv 6 0 1 0 yv 7 7 Tv : Translation by an offset vector v (xv , yv , zv ) D 6 4 0 0 1 zv 5 0 0 0 1 2 1 0 0 6 0 cos sin R™ (x): Rotation by an angle about the x-axis D 6 4 0 sin cos 0 0 0 2 cos 0 sin 6 0 1 0 R™ (y): Rotation by an angle about the y-axis D 6 4 sin 0 cos 0 0 0 2 cos sin 0 6 sin cos 0 R™ (z): Rotation by an angle about the z-axis D 6 4 0 0 1 0 0 0 2 3 xk 0 0 0 6 0 yk 0 0 7 7 Sk : Scaling by factors kx , ky , kz D 6 4 0 0 zk 0 5 0 0 0 1

a

b

1 3 0 07 7 05 1

c 1

v

yk

q zk z

3 0 07 7 05 1 3 0 07 7 05

x

z

x

1

z

xk 1

x

Fig. 2.9 Examples showing transformations of (a) a translation by an offset vector v (b) a rotation about the x-axis by an angle and (c) scaling by factors kx , ky , kz

2.7 Affine Combinations

19

A linear transformation followed by a translation is called an affine transform. A general transformation can be given in matrix form as follows: 3 2 x0 p a00 6 y 0 p 7 6 a10 6 7 6 4 z0 p 5 D 4 a20 0 1 2

a01 a11 a21 0

a02 a12 a22 0

32 3 xp a03 6 yp 7 a13 7 76 7 a23 5 4 zp 5 1

(2.39)

1

In the above equation, the matrix elements aij ’s are all constants. (a03 , a13 , a23 ) denote the translation components, and (xp , yp , zp , 1) the point on which the transformation is applied. The translation parameters do not have any effect on a vector (xv , yv , zv , 0). Under an affine transformation, line segments transform into line segments, and parallel lines transform into parallel lines. A fixed point of a transformation is a point that remains invariant under that transformation. For example, every point along the x-axis is a fixed point for the transformation R™ (x). Similarly, the origin is a fixed point for the scale transformation. The most general rotation of an object with the origin as a fixed point, is the rotation by an angle about an arbitrary vector v D (xv , yv , zv , 0) passing through the origin. The matrix for this transformation is given below. 2

xv yv A zv B xv zv A C yv B xv2 A C C 6 xv yv A C zv B yv2 A C C yv zv A xv B R™ .v/ D 6 4 xv zv A yv B yv zv A C xv B z2v A C C 0 0 0

3 0 07 7 05

(2.40)

1

where A D (1cos™), B D sin™, and C D cos™. A rotation about an axis parallel to the x-axis, with an arbitrary fixed point P, can be obtained by first applying a translation Tp from P to the origin, a rotation R™ (x) with origin as the fixed point, and finally a translation Tp back to the original position P. In matrix form, we write the composite transformation as Tp R™ (x)Tp 1 . Here T1 denotes the inverse of the transformation T. For a translation, the inverse of Tp is Tp ; and for a rotation, the inverse of R™ (v) is R™ (v). A transformation of the form TRT-1 is called the conjugate of R. We have just seen a few examples of affine transformations that are commonly used for generating new points by transforming existing ones. We could also combine the coordinates of a set of points using a linear equation to obtain a new point. Such interpolation methods are discussed in the next section.

2.7 Affine Combinations A linear combination of a set of points Pi (i D 1,2, : : : n) produces a new point Q as shown below: QD

n X i D1

wi Pi

(2.41)

20

2 Mathematical Preliminaries

Fig. 2.10 (a) Linear interpolation and (b) trigonometric interpolation between two points

where the coefficients (weights) wi are constants. If the weights satisfy the condition n X

wi D 1:;

(2.42)

i D1

then Eq. 2.41 gives an affine combination of points. Additionally, if wi 0, for all i, then wi ’s form a partition of unity, and Eq. 2.41 is said to give a convex combination of points. As a special case, when n D 2, we get the formula for linear interpolation between two points P1 and P2 : Q D .1 t/P1 C t P2 ;

0 t 1:

(2.43)

An interesting variation of the above equation can be derived by expressing the parameter t as a function of an angle ’, given by t D cos2 ’. Then the coefficient (1 t) becomes sin2 ’, and Eq. 2.43 takes the form Q D sin2 ’ P1 C cos2 ’ P2 . However, this trigonometric interpolation formula gives a non-uniform distribution of points on the line when ’ is varied from 0ı to 90ı in equal steps. A comparison of linear and trigonometric interpolations is given in Fig. 2.10. In Fig. 2.10a, the parameter t is varied uniformly in the range [0–1] in steps of 0.1, and in Fig. 2.10b, the angle ’ is varied uniformly in the range [0–90] in steps of 9ı . Higher order interpolation between points is discussed in Chap. 7 (Box 2.4).

Box 2.4 Bernstein Polynomials Given a positive integer value n, we can construct n C 1 polynomials of degree n of a parameter t as follows: n .1 t/ni t i ; ˇi;n .t/ D i

i D 0; 1; 2; : : : ; n:

2.7 Affine Combinations

21

These polynomials form a partition of unity, i.e.,

n P i D0

ˇi;n .t/ D 1:

Therefore, they can be used to generate convex combinations of points. Given n C 1 points Pi , i D 0, : : : ,n, we define a point Q(t) as Q.t/ D

n X

ˇi;n .t/ Pi

i D0

As the parameter t is varied from 0 to 1, we get a continuous parametric curve called the Bezier curve. The equations for n D 1, 2, 3 are given below. First degree (linear): Q(t) D (1t) P0 C t P1 Second degree (quadratic) : Q(t) D (1t)2 P0 C 2(1t)t P1 C t2 P2 Third degree (cubic) : Q(t) D (1t)3 P0 C 3(1t)2 t P1 C 3(1t)t2 P2 C t3 P3

Fig. 2.11 A bilinear interpolation scheme first interpolates along the edges to get the values at A and B, and then uses another linear interpolation along the line AB to get the value at Q

P3 A

Q

P1

B

P2

Given a triangle with vertices P1 , P2 and P3 , we can perform a bilinear interpolation between the values defined at the vertices to get the interpolated value at an interior point Q (Fig. 2.11). Using this scheme, we can compute the colour value at any point inside a triangle, given the colour values at the vertices. A scanline parallel to the base of the triangle sweeps the plane and generates the values of A and B using the linear interpolation equation in Eq. 2.43 with the same parameter t. Another linear interpolation between of A and B with a parameter s gives the value of Q. Thus we get Q D .1 s/ f.1 t/P1 C tP3 g C s f.1 t/P2 C tP3 g ;

0 s; t 1: (2.44)

The above equation could be simplified into a simple convex combination of vertex points as Q D .1 k1 k2 /P1 C k1 P2 C k2 P3 ;

0 k1 ; k2 ; k1 C k2 1;

(2.45)

where k1 D s(1t) and k2 D t. The bilinear interpolation of vertex coordinates shown above can be generalized to interpolate any quantity or attribute inside a triangle,

22

2 Mathematical Preliminaries

given its values at the vertices. Examples of such vertex attributes are colour, texture coordinates and normal vectors. In the next section, we will consider another closely related interpolation method for triangles.

2.8 Barycentric Coordinates The barycentre of a rigid body is its centre of mass. For a triangle, the barycentre is its centroid. Given vertices P1 , P2 , P3 of a triangle, the centroid C can be easily computed as the average of the vertex coordinates (P1 C P2 C P3 )/3. Thus C can be represented as a convex combination of the vertex points. Indeed, Eq. 2.45 has just shown that any point Q inside the triangle could be expressed as a convex combination of vertices. If we re-write Eq. 2.45 as Q D 1 P1 C 2 P2 C 3 P3;

0 1 ; 2 ; 3 1;

1 C 2 C 3 D 1; (2.46)

then the point Q is uniquely specified by a new set of coordinates (1 , 2 , 3 ) defined by P1 , P2 , and P3 . This local coordinate system is called the barycentric coordinates for the triangle. Barycentric coordinates are also sometimes referred to as trilinear coordinates. From Eq. 2.46 we see that the vertices themselves have barycentric coordinates given by P1 D .1; 0; 0/ P2 D .0; 1; 0/ P3 D .0; 0; 1/

(2.47)

As seen earlier, the centroid C has barycentric coordinates (1/3, 1/3, 1/3). The barycentric coordinates of a point Q with respect to P1 , P2 , P3 have a geometrical interpretation as the ratios of the areas of triangles QP2 P3 , QP3 P1 , QP1 P2 to the area of the triangle P1 P2 P3 . In the following equations, the symbol denotes the signed area of a triangle: 1 D

QP2 P3 ; P1 P2 P3

2 D

QP3 P1 ; P1 P2 P3

3 D

QP1 P2 P1 P2 P3

(2.48)

The barycentric coordinates given in Eq. 2.48 are unique for every point on the plane of the triangle. They can be directly used to get the interpolated value of any quantity defined at the vertices of the triangle. If fP1 , fP2 , fP3 denote the values of some attribute associated with the vertices, then the interpolated value at Q is given by fQ D 1 fP1 C 2 fP 2 C 3 fP 3 :

(2.49)

2.8 Barycentric Coordinates

23

R= l1S1+l2S2+l3S3

P3

S3 R

Q = (l1, l2, l3)

S2 P1

P2

S1

Fig. 2.12 A one-to-one mapping of points from one triangle to another can be obtained using barycentric coordinates

Using barycentric coordinates we can establish a one-to-one mapping of points from within one triangle to another. For any given interior point Q of the first triangle, we compute the barycentric coordinates. The linear combination of the vertices of the second triangle with the barycentric coordinates of Q gives the coordinates of the corresponding point R inside the second triangle (Fig. 2.12). We can use this mapping to transfer values from the interior of the first triangle to the second. As an immediate application of this transfer, we can map an image (or texture) from one triangle to another. In a simplified two-dimensional case where P1 D (x1 , y1 ), P2 D (x2 , y2 ), P3 D (x3 , y3 ), Q D (xq , yq ), the expressions for the barycentric coordinates of Q given in Eq. 2.48 assume the following form: 1 D

xq .y2 y3 / C x2 .y3 yq / C x3 .yq y2 / x1 .y2 y3 / C x2 .y3 y1 / C x3 .y1 y2 /

2 D

xq .y3 y1 / C x3 .y1 yq / C x1 .yq y3 / x1 .y2 y3 / C x2 .y3 y1 / C x3 .y1 y2 /

3 D

xq .y1 y2 / C x1 .y2 yq / C x2 .yq y1 / x1 .y2 y3 / C x2 .y3 y1 / C x3 .y1 y2 /

(2.50)

If any of the above quantities is negative, then the point Q lies outside the triangle P1 P2 P3. Thus barycentric coordinates find applications in point inclusion tests. In a general three-dimensional case, however, the area of a triangle computed using Eq. 2.3 would always be positive, and correspondingly the area ratios in Eq. 2.48 would also be positive. As previously discussed in Sect. 2.2, the computation of signed areas of triangles requires a view vector w. Since we need this vector to be fixed with respect to every triangle in Eq. 2.48, we can conveniently choose w D (P2 P1 ) (P3 P1 ). Now the barycentric coordinates 1 , 2 and 3 in Eq. 2.48 can be computed by applying the formula in Eq. 2.8 to each of the triangles QP2 P3 , QP3 P1 , QP1 P2 and P1 P2 P3 . If the conditions œ1 C œ2 C œ3 D 1, 0 1 , 2 , 3 1

24

2 Mathematical Preliminaries

are met, then Q lies on the plane defined by the points P1 , P2 , P3 , and also lies within the triangle P1 P2 P3 . Note that in the most general case, the point Q need not be on the plane of the triangle. Hence we require the additional condition that the sum of barycentric coordinates equals 1 to ensure that the points are coplanar. Barycentric coordinates are also useful for finding the centre of a circle that passes through three non-collinear points, P, Q, R in three dimensions. Denoting the vectors along the sides of the triangle by a D QP, b D RQ, and c D PR, the barycentric coordinates of the centre of the circle are 1 D

jbj2 .c a/ 2ja bj2

2 D

jcj2 .a b/ 2ja bj2

3 D

jaj2 .b c/ 2ja bj2

(2.51)

The centre of the circle is then given by the following linear combination of the three points: C D 1 P C 2 Q C 3 R:

(2.52)

In the following section, we will look at the application of vectors in the PhongBlinn illumination model used for lighting calculations in the OpenGL pipeline.

2.9 Basic Lighting The hardware accelerated lighting model that is traditionally used in Computer Graphics applications is based on Phong-Blinn approximation for an omnidirectional point-light source. A local illumination model that does not account for complex effects such as reflections, refractions, shadows and indirect illumination is found to be generally adequate for a majority of graphics applications. In this model, light-material interaction is simply modelled using a component-wise multiplication of material colour and light colour. We can represent colour by a vector comprising of red, green and blue components as c D (r, g, b, 0). This vector model can be further generalized by replacing the fourth component by k that represents the transparency (or opacity) term which can take non-zero values. In the discussion that follows, ma , md , ms denote respectively the ambient, diffuse and specular components of material colour, and Ia , Id , Is the corresponding components of the light source. Each of these colour components is typically a 3-tuple consisting of red, green and blue values. For notational convenience, we represent ma by

2.9 Basic Lighting

25

Fig. 2.13 Important vectors and angles between them, used in lighting calculations

h

n b

r Reflection Vector

s Light Vector

q

q P

f

v View Vector

Surface Element

the vector (rma , gma , bma ), Ia by the vector (ria , gia , bia ), and so on. The ambient light-material interaction is then modelled by the component-wise vector product ma ˝ I a D .rma ri a ; gma gi a ; bma bi a /

(2.53)

Figure 2.13 shows the geometry of unit vectors used for computing diffuse and specular reflections from a surface. From a point P on a surface, s denotes the unit vector towards the light source, n the unit surface normal vector, and v the unit vector towards the viewer. The perceived intensity of reflection at the viewer’s position varies with changes in the angles between these vectors. The variations in diffuse and specular reflections are represented by multiplicative factors kd and ks respectively. According to the Lambertian reflectance model, the intensity of diffuse reflection from a surface is uniform in all directions, and varies as the cosine of the angle between the light source vector s and the surface normal vector n, and is therefore proportional to s•n. If the angle between the two vectors is greater than 90ı , the normal vector faces away from the light source vector, and the surface is in shadow. In such a situation, the value of kd must be set to 0. We therefore have the following view-independent factor for the diffuse term: kd D max.s n; 0/

(2.54)

The specular reflection factor ks is computed as a function of the cosine of the angle between the direction of unit specular reflection r given by Eq. 2.5 and the unit view vector v, with an exponent f known as the shininess term or the Phong’s constant. The exponent is useful in controlling the overall brightness and the concentration of the specular highlight. ks D max.cosf ; 0/ D max..r v/f ; 0/

(2.55)

The Blinn’s approximation eliminates the need for computing the specular reflection vector using Eq. 2.5 by defining a unit vector h along the direction s C v. This vector is called the half-way vector. If n•h D cosˇ, then equating the angles on either side of h gives Cˇ D ˇC

(2.56)

26

2 Mathematical Preliminaries Light Source Ambient Ia

Specular Is

Diffuse Id

ka ks

+

Net Reflection

kd Ambient Ma

Diffuse Md

Specular Ms

Material Colour

Fig. 2.14 Schematic of the calculations performed in a basic lighting model

From the above equation we find that D 2ˇ. The term r•v in Eq. 2.55 can therefore be replaced with n•h by absorbing the factor 2 in ks . This gives the Blinn’s approximation for ks : ks D max..n h/f ; 0/:

(2.57)

A schematic of the lighting computation using the Phong-Blinn illumination model outlined above is given in Fig. 2.14.

2.10 Summary This chapter reviewed some of the geometrical computations involving points, lines, planes, triangles and curves, that are fundamental to many algorithms in computer graphics. Important concepts such as homogeneous coordinate representation of points, signed angles, signed areas of triangles, and barycentric coordinates were outlined. Equations relating to affine transformations and affine combinations of points were discussed. This chapter also gave the equations for a basic lighting model consisting of ambient, diffuse and specular components of reflection. The concepts presented in this chapter will form the foundation for several methods that will be discussed in subsequent chapters. The next chapter introduces a hierarchical structure that is useful for modelling transformations applied to articulated models and other similar objects containing interconnected parts.

2.11 Supplementary Material for Chap. 2 The section Chapter2/Code on this book’s companion website contains code examples demonstrating the application of concepts discussed in this chapter. A brief description of these programs is given below.

2.11 Supplementary Material for Chap. 2

27

1. Point3.cpp

The Point3 class supports most commonly used operations on points represented using 4-dimensional homogeneous coordinates. The class has the subclass Vec3 that supports vector operations such as dot and cross products, vector magnitude calculation and normalization. The documentation of these classes can be found in Appendix A. 2. Triangle.cpp

The Triangle class provides methods for computing area, surface normal vector, and the barycentric coordinates of a point with respect to a triangle. It also has functions for performing the point inclusion test and bilinear interpolation. The documentation of this class can be found in Appendix A. 3. Matrix.cpp

The Matrix class contains methods for matrix operations (using 44 matrices) such as addition, multiplication, computation of transpose and inverse

28

2 Mathematical Preliminaries

matrices, and transformation of points. The documentation of this class can be found in Appendix A. 4. Interpolate.cpp

The program creates a shape-tween between two user-defined polygonal shapes using simple linear interpolation between corresponding vertices. Use left mouse clicks on the upper left side of the screen to define the first polygonal shape. Similarly, use right mouse clicks on the upper right side of the screen to draw the second polygon. Pressing the space bar creates the shape-tween between the first and the second polygons in the bottom half of the window. 5. Bilinear.cpp

The program uses Eq. 2.45 to obtain a bilinear interpolation of color values at the vertices to fill the interior of a triangle. For comparison, a second similar triangle is rendered using the OpenGL pipeline that uses the Gouraud shading algorithm. The vertex colours are randomly generated every time the space bar is pressed. 6. Bezier2D.cpp

The program uses Bernstein polynomials (Box 2.4) to generate a twodimensional Bezier curve for a set of user-defined control points. Use left mouse clicks on the screen to define a set of control points. The control polygonal line

References

29

is shown in red colour. The Bezier curve for the input points is simultaneously drawn in blue colour. 7. Barycentric.cpp

The program uses barycentric mapping (Fig. 2.12) to map points from one triangle to another. Two triangles are displayed when the program is initiated. Use left mouse clicks inside the left triangle to specify a few points. The points are connected using a polygonal line drawn in magenta colour. The map of these points and the polygonal line connecting them inside the triangle on the right hand side are simultaneously drawn in blue colour.

2.12 Bibliographical Notes Several books on introductory computer graphics provide an outline of concepts discussed in this chapter. Some recent publications that can serve as excellent references are Angel (2008), Hill and Kelley (2007), and McConnell (2006). A number of books give emphasis to the mathematical tools used in computer graphics. Notable in this area are Vince and Vince (2006), Lengyel (2004), Buss (2003), Schneider and Eberly (2003), and Dunn and Parberry (2002). Comninos (2006) gives a comprehensive coverage of topics on vector and matrix algebra, transformations, lighting and shading models. A concise description of homogeneous coordinates and their applications in computer graphics can be found in Vince (2001). Topics in linear algebra and topology that are used in many algorithms in computer graphics are discussed at length in Agoston (2005) and Farin and Hansford (2005).

References Agoston, M. K. (2005). Computer graphics and geometric modeling. London: Springer. Angel, E. (2008). Interactive computer graphics: A top-down approach using OpenGL (5th ed.). Boston/London: Pearson Addison-Wesley.

30

2 Mathematical Preliminaries

Buss, S. R. (2003). 3-D computer graphics: A mathematical introduction with OpenGL. New York: Cambridge University Press. Comninos, P. (2006). Mathematical and computer programming techniques for computer graphics. London: Springer. Dunn, F., & Parberry, I. (2002). 3D math primer for graphics and game development. Plano: Jones & Bartlett Publishers. Farin, G. E., & Hansford, D. (2005). Practical linear algebra: A geometry toolbox. Wellesley: A K Peters. Goldman, R. (1990). Intersection of three planes. In A. S. Glassner (Ed.), Graphics gems (Vol. I, p. 305). San Diego: Academic Press. Hill, F. S., & Kelley, S. M. (2007). Computer graphics: Using OpenGL (3rd ed.). Upper Saddle River: Pearson Prentice Hall. Lengyel, E. (2004). Mathematics for 3D game programming and computer graphics (2nd ed.). Hingham/London: Charles River Media/Transatlantic. McConnell, J. J. (2006). Computer graphics: Theory into practice. Boston/London: Jones and Bartlett Publishers. Schneider, P. J., & Eberly, D. H. (2003). Geometric tools for computer graphics. Amsterdam/ London: Morgan Kaufmann. Vince, J. (2001). Essential mathematics for computer graphics fast. London: Springer. Vince, J., & Vince, J. E. (2006). Mathematics for computer graphics (2nd ed.). London: Springer.

Chapter 3

Scene Graphs

Overview A scene graph is a data structure commonly used to represent hierarchical relationships between transformations applied to a set of objects in a three-dimensional scene. It finds applications in a variety of acceleration and rendering algorithms. A scene graph could also be used to organize visual attributes, bounding volumes, and animations as a hierarchy in a collection of objects. In the most general form, any scene related information that can be organized in a hierarchical fashion can be stored in a scene graph. It also provides a convenient way of representing logical groups of objects formed using their spatial positions or attributes. In this chapter, we will outline the fundamental properties of scene graphs, look at some of the implementation aspects and consider a few applications.

3.1 The Basic Structure of a Scene Graph The structure and contents of a scene graph will obviously depend on the type of information it stores, or equivalently, the set of operations it is used for. Let us consider a simple tree structure that contains three types of nodes: 1. The root node of the tree represents the whole collection of objects in a threedimensional scene. We call this node World or Virtual Universe. The root node is a special type of a group node. 2. A group node is an internal node of the tree. It can contain any number of children, and represents a logical grouping of objects. A group node does not store geometrical data, but it can contain some semantic information such as transformations or visibility attributes applied to a group. 3. Every leaf node represents either an object or a part of an object, and maintains the necessary geometrical information in addition to some semantic information. Camera and light sources may also be represented by leaf nodes. R. Mukundan, Advanced Methods in Computer Graphics: With examples in OpenGL, DOI 10.1007/978-1-4471-2340-8 3, © Springer-Verlag London Limited 2012

31

32

3 Scene Graphs

World

Group

Object

Object

Group

Object

Group

Object

Object

Object

Object

Fig. 3.1 An example of a scene graph, where every internal node is a group node and every leaf node is an object node

a

Group-2 Part-3

b

World

Group-1

Part-1

Part-2 Part-1

Group-2

Base

Base Part-2

Part-3

Fig. 3.2 (a) An example of a model consisting of four connected parts that can move relative to each other. (b) A scene graph of the object model

Figure 3.1 shows an example of a tree with all three types of nodes described above. The tree structure of a scene graph allows a property associated with a group node to be inherited by all of its child nodes. For example, a transformation applied to a group node can be considered as also applied to all its children. Similarly, a bounding volume, if attached to a group node, also represents the overall bounding volume for the whole collection of its child nodes. A scene graph is particularly useful for animating a composite object that has several parts which should move as if the parts are all physically connected to each other. A typical example of such an object is an articulated character model. We illustrate the formation of a scene graph using a simple model consisting of four interconnected parts: Base, Part-1, Part-2, and Part-3, as shown in Fig. 3.2.

3.2 Transformation Hierarchy

33

World Link-3

Link-4 Group-1

Link-1 Link-5 Link-2

Group-2

Link-2 Group-3

Link-3 Link-1

Link-4

Group-4

Link-5

Fig. 3.3 A 5-link joint chain and its scene graph

As can be seen from the diagram of the scene graph, the whole model is first subdivided into three logical groups Part-1, Base and a subgroup Group-2 to which Part-2 and Part-3 belong. Shortly we will see how we can assign transformation parameters to the individual nodes of the scene graph in such a way that the parts can rotate relative to each other while at the same time remaining connected as a single animatable object. We now consider a closely related object model, a joint chain consisting of five links as shown in Fig. 3.3. Joint chains similar to the one shown above are commonly found in robotics and articulated models in computer graphics. The scene graph represents a hierarchical subdivision of the model, where at the first level, the whole object belongs to a single group World. At the next level of subdivision we have Link-1 and a subgroup Group-1 that contains the remaining links. Any rotational transformation applied to Group-1 affects all members of that group. It may appear that the group node Group-4 is redundant as it has only one child. However, the node is useful to provide a clear separation between the initial transformations applied to the object in Link-5 in its own coordinate system and the transformations applied relative to Link-4’s frame. We will also later add a camera as an object belonging to Group-4. The transformation hierarchy represented by scene graphs is explored in more detail in the next section.

3.2 Transformation Hierarchy A transformation applied to one part of an object often cascades with the transformations applied to the adjacent interconnected parts. For example, a change in the orientation of Part-2 of the model in Fig. 3.2a also affects Part-3. Such dependencies can be easily converted into hierarchical representations that are suitable for scene graphs. We consider below three examples involving hierarchical transformations: (i) the model of a mechanical part shown in Fig. 3.2, (ii) an articulated character model, and (iii) a small planetary system.

34

3 Scene Graphs d3 Part-3 Part-2 d2 y

q

Part-1 d1 Base a

0

x

Fig. 3.4 A general transformation of the model in Fig. 3.2, showing translational and rotational parameters associated with links. The x and y axes denote the reference frame for the world coordinate system

3.2.1 A Mechanical Part A general two-dimensional transformation of the model in Fig. 3.2a along with the translational and rotational parameters of each link is shown in Fig. 3.4. We will use T(a) to denote a translation by a vector a, and R() to denote an anticlockwise rotation through an angle . Note that the joint angles ı 1 , ı 2 , ı 3 define relative angles of rotations of one part with respect to another. In order to build the transformation hierarchy, we have to consider first the transformation of each link from its own local coordinate frame to the coordinate frame of its group. The sequence in which the transformations are applied is shown in Fig. 3.5. As shown in Fig. 3.5, transformations are applied from the leaf nodes upward to the root of the scene graph. Part-3 is first rotated by an angle ı 3 , and then translated along the length of Part-2 by a vector d3 . This composite transformation has a matrix given by T(d3 )R(ı 3 ). Group-2 now contains Part-2 and the transformed version of Part-3. In other words, both Part-2 and Part-3 have been transformed into the coordinate space of Group-2. It should be noted here that any rotational transformation of Part-2 is always applied to Group-2. The transformation matrix T(d2 )R(ı 2 ), effectively converts the points from the coordinate system of Group-2 to that of its parent group, Group-1. Figure 3.6 shows the scene graph with the transformation matrices added to the tree nodes. From the above discussion, we note that every node transformation is defined relative to the node’s parent. At a leaf node, a transformation converts vertices from the local coordinate space of an object to its parent’s coordinate space. If an object node has an identity transformation I, it only shows that its parent’s node has the same coordinate reference frame as the object node. This also means that any transformation applied to that node is actually applied to its parent group node. In the above example, transformations applied to the Base are actually applied to Group-1, and they indirectly affect the transformations of each of Group-1’s child nodes.

3.2 Transformation Hierarchy

35

T(d3)R(d3)

y

y

T(d2)R(d2)

x

T(a)R(q)

Part-3 I

y

x

d3

Group-2 y x

y

y

d2

Part-2

a x

x

World

Group-1

Base d1

x

I T(d1)R(d1) x Part-1

Fig. 3.5 Each moveable component of an object model is transformed from its local coordinate space to its group’s space, and subsequently to the coordinate space of the group’s parent

World

T(a)R(q)

Group-1

Part-1

Base

T(d1)R(d1)

I

Group-2

Part-2 I

T(d2)R(d2)

Part-3 T(d3)R(d3)

Fig. 3.6 Scene graph with transformation matrices attached to nodes

3.2.2 A Simple Character Model We now consider an articulated character model and its scene graph shown in Fig. 3.7. As in the previous example, we can define the translational and rotational transformations for each node, based on the joint position and angle of each link relative to its parent. Vectors v1 : : : v9 denote the offsets of the origin of the links relative to their parent’s local coordinate system in the initial configuration. The vector v0 denotes the position of the base link (Torso) in the world coordinate frame.

36

3 Scene Graphs

World Group-1 T(v0)R(yz) R(yy) R(yx)

I

Torso T(v2)R(q2) Group-2

Head

T(v4)R(q4) Group-3

Left Upper Arm

Right Lower Arm

Right Upper Arm

T(v3)R(q3)

I

T(v5)R(q5)

I

T(v8)R(q8)

T(v6)R(q6) Group-4

Left Lower Arm

T(v1)R(q1)

Left Lower Leg T(v7)R(q7)

Group-5

Left Upper Leg I

Right Lower Leg T(v9)R(q9)

Right Upper Leg I

Fig. 3.7 Scene graph of a basic articulated character model

The angles x , y , z represent a generalized rotation of the whole model in terms of Euler angles defined with respect to the principal axes of the world coordinate system. A detailed description of Euler angle rotations can be found in Sect. 5.4.1. The model can be animated using key-frame sequences for the joint angles 1 .. 9 , and its position and orientation can be controlled using key-frame sequences for v0 , x , y , and z . The transformation hierarchy, if properly defined, ensures that the links stay connected and are rotated only about the joints. Owing to the symmetry of the model, we can also make use of the following relationships among the components (xi , yi , zi ) of translational parameters vi : x2 D x4 I

y2 D y4

x3 D x5 I

y3 D y5

x6 D x8 I

y6 D y8

x7 D x9 I

y7 D y9

(3.1)

3.2.3 A Planetary System As the third example, we consider a simple planetary system consisting of the Sun, the Earth and the Moon. The translational and rotational parameters used in modelling the system are shown in Fig. 3.8. The rotation angles E , M represent the spin of the Earth and the Moon respectively about vertical axes, E denotes the revolution of the Earth-Moon

3.2 Transformation Hierarchy

37

qM

qE

fE Sun Moon

Earth fM

dE

dM

Fig. 3.8 A simple planetary system showing the translational and rotational parameters used for the construction of its scene graph

World

Sun

Group

R(fE)T(dE)

I Earth R(qE)

Moon R(fE)T(dM)R(qM)

Fig. 3.9 Scene graph of the planetary system

system around the Sun, and M the revolution of the Moon around the Earth. The scene graph for this system is shown in Fig. 3.9. One notable difference between the planetary system example and the previous ones is the form of transformation matrices applied to nodes. Most of the transformations applied in a hierarchical fashion have a general form T(v)R(), which is a rotation followed by a translation. In simple implementations, the structure of nodes is often designed to accept only transformations of the form T(v)R() or I. Scene graphs where transformations at internal nodes have one of the forms I, T(v), R(), or T(v)R() are said to be in the standard form. The example given in Fig. 3.9 is an exception to this rule. However, this scene graph can be easily converted to the standard form with the addition of a group node as shown in Fig. 3.10. The equivalence of the scene graphs in Figs. 3.9 and 3.10 can be verified by obtaining the combined final transformation matrices applied to the leaf nodes. In a scene graph, transformations are combined using a recursive procedure starting at the root node, accumulating transformations at internal nodes and ending at object nodes. This process will be explained in detail in the next section.

38

3 Scene Graphs

World

Group-1

Sun

R(fE)

I Earth

Group-2

T(dE)R(fM)

T(dE)R(qE) Moon T(dM)R(qM)

Fig. 3.10 The scene graph in Fig. 3.9 converted to the standard form

3.3 Relative Transformations The transformation of one node relative to another can be readily obtained from a scene graph. The model transformation matrix of an object gives the composite transformation that converts points from the local coordinate space of the object to the world coordinate space. In a scene graph, this is the transformation of the object node relative to the root (the world node). The composite matrix can be obtained by collecting all matrices along the path from the root node to the leaf node representing the object. At each node, the matrix is post-multiplied by the transformation matrix of that node. The process is illustrated in Fig. 3.11, where node transformation matrices are denoted by letters A..G. The model transformation matrix of the object node in the figure is ABCDE. Leaf nodes can also be used to represent fictitious objects such as light sources and camera. In Fig 3.11, the transformation from the coordinate system of the camera to world coordinates is given by AFG. The inverse of this matrix, (AFG)1 , transforms a point from world space to camera space. This matrix is called the view matrix. The combined model-view matrix that transforms the object’s local coordinates to camera space is therefore given by (AFG)1 ABCDE, or equivalently, G1 F1 BCDE. An upward tree traversal from a leaf node to root can be quickly performed if every node has a pointer to its parent. On the other hand, a downward traversal would typically require a recursive algorithm similar to the depth-first search method. The above example can be generalized to a procedure for finding the transformation from one object’s local coordinate frame to another’s. If we require the transformation from Object-1 (source) to Object-2 (target) in a scene graph, we have to first find the Lowest Common Ancestor (LCA) of both the object nodes. Let the transformation matrix of this common ancestor be denoted by M (Fig. 3.12). Let S1 : : : Sm denote the transformations of nodes starting from the child of LCA

3.3 Relative Transformations

39

World Model Transformation Matrix

A

View Matrix F

B

Camera

C

G

D

Object

E

Fig. 3.11 Computation of the model transformation matrix of an object represented by a leaf node in a scene graph

LCA

X

S1

M

Y

T1

Object-2 Object-1

Tn

Sm

Fig. 3.12 Representing Object-1’s coordinates relative to Object-2’s local reference frame requires the computation of the Lowest Common Ancestor (LCA) of both the nodes

towards Object-1, and T1 ..Tn the transformations towards Object-2 as shown in Fig. 3.12. The composite transformation from the source’s frame to the target’s frame is given by the matrix Tn 1 ..T1 1 S1 ..Sm . Note that this matrix product does not involve the transformation M of the LCA or any of its ancestors. There are several well-known algorithms to compute the Lowest Common Ancestor of two nodes in a tree. A simple method uses two lists of nodes visited in sequential upward traversals of the tree from the two nodes towards the root. The last item of both lists would be the world node. Corresponding entries in the

40

3 Scene Graphs Object-1

X

LCA

World

List-1: Comparison ends (X≠Y)

Comparison starts here

List-2:

Object-2

Y

LCA

World

Fig. 3.13 An algorithm for finding the Lowest Common Ancestor

lists are compared for equality, starting from the last item towards the beginning of each list. The process of comparison stops when the list entries are different. The previous matched entry in the lists gives the reference to the Lowest Common Ancestor (Fig. 3.13).

3.4 Bounding Volume Hierarchy Bounding volumes of objects are used for fast collision detection and also in acceleration algorithms such as view frustum culling. Bounding volumes can be computed for different moving parts of an object and then combined in a hierarchical manner to obtain the overall bounding volume (Fig. 3.14). The geometric parameters defining a bounding volume can be stored in a scene graph node, and computed on the fly whenever a transformation is applied to the vertices. Commonly used bounding volumes are axis-aligned bounding boxes (AABB), oriented bounding boxes (OBB), spheres, discrete oriented polytopes, and convex hulls. Each bounding volume has certain advantages and limitations over others, and is suitable for a specific set of applications. An AABB can be computed and represented using six parameters that define the minimum and maximum values of x, y, and z coordinates of points it encloses. However, these parameters will have to be recomputed every time an object is rotated. On the other hand, OBBs and spheres are rotation invariant. In this chapter, examples are provided using AABBs and spheres only. Other types of bounding volumes and their computational aspects are discussed in detail in Chap. 9. Since the bounding volume parameters depend on the transformed object coordinates, bounding volume updates can be performed only after applying the transformations. Unlike transformations, this process starts at the nodes containing object primitives, and the bounding volume parameters of group nodes are updated

3.4 Bounding Volume Hierarchy

41

Fig. 3.14 Two-dimensional bounding volume hierarchies for the model in Fig. 3.2, using axisaligned rectangles (top row) and circles (bottom row)

a

b

c

Fig. 3.15 (a) Bounding circles of two objects. (b) Combined bounding circle formed using the parameters of the two component bounding circles. (c) The minimal bounding circle

based on the computed values at the child nodes. It is therefore often desirable that the parameters defining a bounding volume stored at a group node can be computed based on the bounding volume parameters of its child nodes. It should also be noted here that such a computation may not always yield a minimal bounding volume. For example, the bounding sphere computed as the union of two bounding spheres may not necessarily be the minimal bounding sphere for the union of points within those spheres. A two-dimensional equivalent of this case is shown in Fig. 3.15, using bounding circles of two objects. We discuss below the process of updating the bounding volume parameters (using AABBs and spheres as examples) at a group node based on the updated parameters of its child nodes. If there are n child nodes, we combine the volumes of

42

3 Scene Graphs

Box 3.1 Bounding Volumes Given a set of mesh vertices with coordinates fxi , yi , zi g, i D 0 : : : N1, the bounding volume parameters for AABB and sphere are computed as follows: Axis Aligned Bounding Box (AABB): fxmin , ymin , zmin , xmax , ymax , zmax g xmin D mini .xi / ; xmax D maxi .xi / ymin D mini .yi / ; ymax D maxi .yi / zmin D mini .zi / ; zmax D maxi .zi / Sphere: fu, v, w, rg Computation of bounding sphere using the geometric centre of points: u D; N1

NP 1 i D0

xi ;

di D .xi u/2 C .yi v/2 C .zi w/2 ;

i D 0 : : : N 1: NP 1 v D N1 yi ; wD

1 N

i D0 NP 1 i D0

zi ;

rD

p maxi .di /

Computation of bounding sphere using AABB of points: 1 1 1 .xmin C xmax / ; v D .ymin C ymax / ; w D .zmin C zmax / 2 2 2 q 1 .xmax xmin /2 C .ymax y min /2 C .zmax zmin /2 rD 2

uD

two children at a time and obtain the final bounding volume of the parent, in n1 steps. Given two AABBs with parameters fxmin1 , ymin1 , zmin1 , xmax1 , ymax1 , zmax1 g and fxmin2 , ymin2 , zmin2 , xmax2 , ymax2 , zmax2 g, the combined volume has parameters fmin(xmin1 , xmin2 ), min(ymin1 , ymin2 ), min(zmin1 , zmin2 ), max(xmax1 , xmax2 ), max(ymax1 , ymax2 ), max(zmax1 , zmax2 )g (Box 3.1). In the case of spheres, let the parameters of the two volumes be given by fu1 , v1 , w1 , r1 g and fu2 , v2 , w2 , r2 g. The required parameters of the combined sphere are denoted as fuc , vc , wc , rc g. First we compute the distance between the centres: q d D

.u2 u1 /2 C .v2 v1 /2 C .w2 w1 /2

(3.2)

If d jr1 r2 j, then one of the spheres is inside the other. The combined sphere in this case is the same as the larger among the two spheres. If d > jr1 r2 j, the spheres either overlap or are disjoint. For this configuration, we compute the radius and the centre of the combined sphere as follows:

3.5 Sample Implementation

1 .d C r1 C r2 / 2 1 .d r1 C r2 / .u2 u1 / uc D u1 C 2d 1 vc D v1 C .d r1 C r2 / .v2 v1 / 2d 1 .d r1 C r2 / .w2 w1 / wc D w1 C 2d

43

rc D

(3.3)

A detailed description of different types of bounding volumes, their computation and intersection tests is given later in Sect. 9.1.

3.5 Sample Implementation In this section, we will discuss the design of a set of classes that implement the functionality of a scene graph with transformation matrices attached to its nodes. Internal nodes that can store a list of children, and also a transformation matrix, are represented by the class GroupNode. All transformation matrices are assumed to have the general form given by T(v)R(). The properties of leaf nodes are specified by three classes: ObjectNode that can represent a three-dimensional object, CameraNode that represents the camera, and LightNode that represents a light source. These three classes are derived from GroupNode so that we can store all child nodes (including group nodes and object nodes) with the same type, and also use polymorphic functions to implement tree traversal algorithms.

3.5.1 Group Node The declarations of attributes and functions of GroupNode can be found in Listing 3.1 below. The primary functions associated with a group node include adding and removing children, and setting the transformation parameters. We use the List container of the Standard Template Library (STL) for storing references to the child nodes. The data members angleX, angleY, angleZ specify the Euler angles of rotation about the principal axes of the group’s coordinate frame. Similarly tx, ty, tz denote the components of the translation vector along the principal axes directions. Together, these attributes define the composite transformation for the group node in the form T(v) Rz ( z ) Ry ( y )Rx ( x ), where v is the translation vector, and s denote Euler angles. The function render() is called on the root node to render the scene.

44

3 Scene Graphs Listing 3.1 Class definition for a group node

3.5.2 Object Node The class definition for an object node must cater to the requirements of defining and storing three-dimensional object models. Listing 3.2 gives the declarations of important attributes and functions of the class. To simplify the implementation, we use only the built-in objects provided by the GL Utility Toolkit (GLUT) of the OpenGL API. These objects are assigned numbers using the enumerated type ObjType. When an object is initially defined using the setObject() function, it may also be optionally scaled using parameters scaleX, scaleY and scaleZ. These parameters are used to set the values of the corresponding data members of the class. An object may also be given a material colour using the function setColor(). A scene is rendered by calling the function render() of the GroupNode class on an instance that represents the scene graph’s root. This function in turn calls the polymorphic function draw()which is declared as virtual in GroupNode. The implementation of the function in ObjectNode will call the necessary OpenGL functions to apply the transformations and to draw the object.

3.5 Sample Implementation

45

Listing 3.2 Class definition for an object node

3.5.3 Camera Node Any three-dimensional scene is assumed to have an active camera that contains information about the projective transformation used while rendering the scene. The camera also provides the view matrix needed for the transformation of vertices to the eye coordinate space. A camera can be added to a scene graph as a special type of object node. Listing 3.3 gives the class definition for the camera node. Since only one instance of the camera is used in a scene at any point in time, the class cameraNode is defined as a singleton class. It has a private constructor, and the static instance is made available to a program using the function getInstance(). The frustum parameters are specified by an application by calling the function perspective(). The function projection() uses these parameters to set up the projection matrix, and is called by render() of the GroupNode class. The view transformation matrix is constructed by the function viewTransform() by traversing the tree along the path from the camera node to the root node (Fig. 3.11). The class does not store any drawable object, and therefore draw() has an empty function body.

3.5.4 Light Node The LightNode class as defined in Listing 3.4 has a simple structure containing no public functions other than the constructor. The constructor accepts a single integer between 0 and 7 as the argument which directly represents one of the OpenGL light

46

3 Scene Graphs Listing 3.3 Class definition for a camera node

Listing 3.4 Class definition for a light node

sources GL LIGHT0, : : : ,GL LIGHT7. In OpenGL, light sources are transformed like any other point. The function draw() defines the initial position of the light source at (0,0,0), and transforms it exactly like its counterpart in ObjectNode. The class does not store or set any other light or material properties. They can be set by the application by directly calling the appropriate OpenGL functions. The same applies to setting OpenGL states such as enabling lighting, selecting two sided lighting, enabling colour material, and so on. The sample implementation of a scene graph discussed above concatenates only transformation matrices along different paths from the root node to the leaf nodes. The hierarchical structure of a scene graph allows several other attributes to be propagated from an internal node to object nodes through various branches. One such attribute is the visibility of a node. If a node’s visibility attribute is set to false, then the visibility attribute of every node in that sub-tree can also be implicitly set to false by using a logical AND operation with the values from the parent nodes. Thus

3.6 First-Person View

47

an object node will not be rendered if any of its ancestors has a visibility attribute set to false. A similar attribute that can be attached to the nodes is transparency. The transparency values can be multiplied together along every path from the root node, to determine the net transparency of objects stored in the leaf nodes.

3.6 First-Person View The design of the camera node as outlined in the previous section permits a highly flexible implementation of a scene graph, since the only static instance of the class can be obtained anywhere by calling the getInstance() function. The camera node need not even be a part of the scene graph, if the camera is meant to be in a fixed location with respect to the scene. In this case, the transformations defined for the camera node specify the position and the orientation of the camera with respect to the origin of the world coordinate frame. These transformations will be directly used to obtain the view matrix for the whole scene. Often you will require the first-person view of a scene with the camera placed on a moving object. For the articulated character model in Fig. 3.7, the first-person view is provided when the camera is attached to the head. This is done by first applying transformations to the camera node so that it points to the right direction in the coordinate frame of the object node to which it should be attached. In the scene graph, the object node is replaced by a new group node. Both the camera node and the object node are attached to the new group node as its children. Figure 3.16 shows the reference frame (xe , ye , ze ) of the camera and the coordinate frame (x, y, z) of the head of the character model. The camera initially points towards –ze direction. It is rotated about the y-axis by 180ı to point towards the head direction. This transformation is represented by the matrix R( ). Figure 3.16 also shows the modified portion of the scene graph in Fig. 3.7 with the addition of a new group node and the camera node. Now consider the 5-link joint chain shown in Fig. 3.3. Robotic arms such as this can be found in autonomous systems for inspection, welding and painting. The arm is driven by feeding joint angles to the controllers. Some constraints may be applied to the joint angles based on the application requirements. For example, a robotic arm for welding or painting may require the end effector (denoted by Link-5 in Fig. 3.3) to be kept in a horizontal position. It may also be required to have a camera attached to the end effector to obtain a clear perspective of the surrounding scene from its viewpoint. The graphical rendering of the scene as viewed from the position of Link-5 can be obtained by adding the camera node to the group node Group-4 as shown in Fig. 3.17. From the previous examples, we have seen that the first step in the process of attaching a camera to an object node is to determine the transformation R( ) necessary to appropriately orient the camera in the local coordinate frame of the object. In the example in Fig. 3.17, this composite transformation comprises of two

48

3 Scene Graphs

c World

a

Group-1 T(v0)R(yz)R(yy)R(yx)

b

ye

y

View axis

New Group

f

xe

x

ze

Camera

z

T(v1)R(q1)

Head

R(f)

I

Fig. 3.16 (a) Camera coordinate system. (b) A 3D object “Head” in its local coordinate frame. (c) The modified portion of the scene graph in Fig. 3.7, with the camera node attached

a

b

y

View Axis

c

Group-4 x z

ye xe

ze

Camera

Link-5

R(f)

I

Fig. 3.17 (a) Local coordinate frame of a link of the joint chain in Fig. 3.3. (b) The desired orientation of the camera frame relative to the frame of the link. (c) Addition of the camera node to the scene graph in Fig. 3.3

rotations: a rotation of 90ı about the x-axis followed by another rotation of 90ı about the y-axis. The transformation functions given in Listing 3.1 allow us to define such rotations. It is also important to note that when a new group node is formed with the camera node and the object node as its children, transformations that were previously applied to the object node should now be applied to the camera as well. Therefore, the transformation matrix that was attached to the object node must now be transferred to the common group node. This would often leave the object node with the identity matrix as shown in Fig. 3.17.

3.8 Supplementary Material for Chap. 3

49

3.7 Summary Scene graphs are powerful data structures that can be used for hierarchical representations of transformations, bounding volumes and other visual attributes of groups of objects in a scene. This chapter showed the application of scene graphs in defining the transformations of interconnected systems. Robotic manipulator arms and articulated character models are examples of such systems containing one or more joint chains. Using a scene graph, the relative transformation of one object with respect to another can be easily computed. Relative transformations are useful for displaying billboards and first person views. This chapter also introduced the definition of a scene graph in the standard form. An object oriented framework for a scene graph was presented and some of the key implementation aspects were discussed. The next chapter will show that scene graphs play an important role in skeletal animation. Skeletal structures and the associated hierarchical transformations used in vertex skinning algorithms fit perfectly well with the scene graph model.

3.8 Supplementary Material for Chap. 3 The folder Chapter3/Code on the companion website contains code examples demonstrating the application of the scene graph class in the modelling and rendering of simple three-dimensional scenes. A brief description of these programs is given below. 1. GroupNode.cpp

These are the header and implementation files for a scene graph class as discussed in Sect. 3.5. The documentation of methods in this class can be found in Appendix B.

50

3 Scene Graphs

2. Scene3D.cpp

This program uses a scene graph to model a scene consisting of four different stationary objects and demonstrates the use of the classes discussed in Sect. 3.5. The scene graph has a simple structure consisting of the World node and four object nodes. The camera node is not attached to the scene graph and is independently transformed to simulate camera motion along a circular path around the group of objects. The light source is kept fixed in the middle of the scene, at its default position (0, 0, 0). 3. Planet.cpp

This program uses the scene graph in Fig. 3.10, to model the planetary system in Fig. 3.8. The angles of revolution of the Moon around the Earth, and the joint Earth-Moon system around the Sun are continuously updated to generate an animation sequence. The light source is kept fixed at the location of the Sun. 4. Link5.cpp

This program uses the scene graph model given in Fig. 3.3, to construct an animated 5-link robotic arm. The joint angles are read in from the file JointAngles.txt. The arm moves continuously up and down in front of a vertical coloured wall. The joint angles are defined such that the end effector of the arm is always horizontal. Pressing ‘c’ on the keyboard causes the scene graph to be modified as in Fig. 3.17 to produce the effect of the camera being placed on the end effector. This gives a close-up view of the coloured wall from the perspective of the continuously moving end effector.

3.9 Bibliographical Notes

51

5. GlutMan.cpp

The program GlutMan demonstrates the use of a scene graph in modelling and animating an articulated character model. A scene graph similar to the one given in Fig. 3.7 is used. The values of eight joint angles defining a simple walk sequence are read from the input file WalkCycle.txt and interpolated to generate a continuous animation sequence.

3.9 Bibliographical Notes An excellent introduction to scene graphs and other tools for scene management can be found in Sherrod (2007). The book also deals with the design of data structures and algorithms for similar applications. Angel (2008), McConnell (2006) and McReynolds and Blythe (2005) give an overview of hierarchical modelling techniques and applications using scene graphs. Eberly (2007) contains a chapter on hierarchical scene representations, and provides a detailed description of scene graph operations designed for merging a set of bounding volumes. Support for scene graphs including sophisticated high-level functionalities can be found in graphics APIs. Java-3D provides powerful classes for constructing the nodes of a scene graph that can be used for rendering scenes. Many examples of applications in Java can be found in Davison (2005). The M3G API of Java Micro Edition also contains a versatile collection of methods useful for retained-mode rendering based on scene graphs. These methods incorporate high-level functions for generating key-frame animations on mobile devices. Pulli (2008) provides an excellent coverage of the M3G API and shows the importance of scene graphs in the design of animation sequences. OpenSceneGraph is a versatile high-level 3D graphics toolkit useful for the development of high-end graphics applications based on a full-fledged and powerful scene graph implementation. More information can be found on the website, http://www.openscenegraph.org.

52

3 Scene Graphs

References Angel, E. (2008). Interactive computer graphics: A top-down approach using OpenGL (5th ed.). Boston/London: Pearson Addison-Wesley. Davison, A. (2005). Killer game programming in Java. Beijing/Farnham: O’Reilly. Eberly, D. H. (2007). 3D game engine design: A practical approach to real-time computer graphics (2nd ed.). Amsterdam/London: Morgan Kaufmann. McConnell, J. J. (2006). Computer graphics: Theory into practice. Boston/London: Jones and Bartlett Publishers. McReynolds, T., & Blythe, D. (2005). Advanced graphics programming using OpenGL. Amsterdam/London: Morgan Kaufmann Publishers. Pulli, K. (2008). Mobile 3D graphics: With OpenGL ES and M3G. Amsterdam/London: Elsevier/Morgan Kaufmann Publishers. Sherrod, A. (2007). Data structures and algorithms for game developers (1st ed.). Hingham/Charles River Media/London: Thomson Learning [distributor].

Chapter 4

Skeletal Animation

Overview This chapter discusses concepts such as vertex blending, vertex skinning and keyframing that are fundamental to the animation of articulated character models. Vertex blending is the process of constructing blending surfaces between two different parts that move relative to each other, in order to create the appearance of a single deformable object. Vertex blending is useful in the animation of character models constructed by joining together several individual components. Mesh models of animatable characters are often subdivided into groups of vertices that represent moveable body parts. A skeleton is an abstract representation of this form of partitioning of a mesh. Skeletal animation refers to the process of computing the transformations of each segment in the skeleton using joint angles, and mapping them on to mesh vertices. The chapter discusses various stages in skeletal animation, describes the transformations applied to a mesh, and also outlines a scene graph based implementation.

4.1 Articulated Character Models Animated character models can be found in numerous applications of computer graphics, ranging from simple computer games to virtual agents and computer generated feature films. Depending on the application requirements, the character mesh and the animation sequence can have varying levels of complexity. Sophisticated virtual character agents incorporate several forms of articulation including facial expression animation. In this chapter we will look at the basics of human character animation with simple polygonal models and a small number of joint angles. We broadly classify character models into two groups: (i) character models constructed using several objects or “parts” where each object is independently transformed and moved into its respective position within the model, and (ii) single mesh models that are animated by attaching vertices to different transformation R. Mukundan, Advanced Methods in Computer Graphics: With examples in OpenGL, DOI 10.1007/978-1-4471-2340-8 4, © Springer-Verlag London Limited 2012

53

54

4 Skeletal Animation

Fig. 4.1 Character models constructed using (a) several component objects, and (b) a single mesh

groups. An example of each type is shown in Fig. 4.1. The first model, the “Glut Man”, is constructed entirely using scaled and transformed versions of cubes generated using glutSolidCube() or glutWireCube(), hence the name. The second belongs to the more commonly found class of mesh models. In the case of the model constructed using individual parts, each component is first created in its own local coordinate space. A series of transformations is then applied to it based on where in a joint chain that component appears. This process, which is very similar to what we saw in the previous chapter (Fig. 3.5), is repeated for every part of the model to reshape the character in a required pose. The transformations often have a well-defined hierarchical structure as discussed in the context of scene graphs. Figure 3.7 shows how the main body parts of a simple humanoid model are transformed. A character model defined using a single mesh surface as in Fig. 4.1b requires a completely different set of coordinate transformations, as all mesh vertices are specified in a common reference system. However, we should be able to use the same set of joint angles to animate this model also, producing a similar effect such as a walk cycle. We can indeed construct a “virtual” skeleton consisting of joints and links that has a structure similar to our previous model in Fig. 4.1a. We can then associate the skeleton with the continuous mesh. This association is done by attaching a set of vertices belonging to each body part (e.g., forearm) to the corresponding link of the skeleton. The scene graph based transformations computed using joint angles can now be directly applied to the skeleton. The mesh vertices are transformed using a simple method introduced in Sect. 4.4.2. If a model is made up of several parts as in Fig. 4.1a, where parts move or rotate relative to their neighbours, gaps can appear at joints when the model is animated. The next section addresses this problem, and introduces the method of vertex blending for creating deformable surface patches between parts that move relative to each other.

4.2 Vertex Blending

55

4.2 Vertex Blending When two different mesh objects attached to a common pivot rotate by different angles, certain parts of the surfaces can interpenetrate, and gaps can appear on the opposite side (Fig. 4.2a). Repairing or “re-meshing” an area where two surfaces interpenetrate is a difficult task. Moveable surfaces are therefore often separated by a small distance from each other, so that they do not touch for the allowable range of movement or rotation angles (Fig. 4.2b). A sphere is sometimes placed at rotary joints, as in Fig. 4.2c, to fill the gap. While this approach is suitable for robot-like models, interpolation methods could be used for obtaining a better approximation of blending surfaces between moving parts. The process of creating such in-between surfaces is called vertex blending. Corresponding pairs of points on two moving parts can be joined together to form a triangular or quadrilateral element belonging to the intermediate surface. These elements could be further subdivided using a simple linear interpolation formula (Eq. 2.43) to get a tessellated surface (Fig. 4.3a). We discuss below higher order interpolation methods for generating blending surfaces (Fig. 4.3b). In Chap. 2 (Sect. 2.7) we saw examples of second and higher degree interpolation functions with Bernstein polynomials as basis. We can use cubic Bezier polynomials to generate interpolating curves between moving parts with tangential continuity at end points. In Fig. 4.4a, P0 and P3 denote a pair of corresponding points on two moving parts of a character model. Q0 and Q3 are two points on the surfaces that are selected to define the local tangent directions P0 Q0 and P3 Q3 respectively. Using these tangent directions, we can specify two more points, P1 and P2 , as P1 DP0 C ˛.P0 Q0 / P2 DP3 C ˛.P3 Q3 /

(4.1)

where ˛ is a positive quantity used to increase or decrease the length of the tangent vectors P1 P0 and P2 P3 . Points on the interpolating Bezier curve are generated

Fig. 4.2 (a) Moving parts of an animated model can interpenetrate and form gaps at joints. (b) Links can be separated by a short distance to avoid surface intersections. (c) A sphere is sometimes attached to a rotary joint to fill the gap between two moving parts

56

4 Skeletal Animation

Fig. 4.3 Generation of blending surfaces using (a) linear interpolation and (b) Hermite interpolation

Fig. 4.4 Generation of a blending surface using (a) Bezier interpolation and (b) Hermite interpolation

a

b Q3

Q3 P3

P3

P2 P1 P0

P0 Q0

Q0

using the parametric equation (see Box 2.4, Sect. 2.7) Q.t/ D .1 t/3 P0 C 3.1 t/2 tP1 C 3.1 t/t 2 P2 C t 3 P3 ;

0 t 1: (4.2)

Substituting the expressions from Eq. 4.1 in the above equation gives Q.t/ D.1 3t 2 C 2t 3 / P0 C 3.1 t/2 t ˛.P0 Q0 / C 3.1 t/t 2 ˛.P3 Q3 / C .3t 2 2t 3 / P3

(4.3)

4.3 Skeleton and Skin

57

When ˛ is increased, the weight of the tangent vectors on the interpolating curve is increased, and the curve gets closer to the tangents at the end points P0 , P3 . Care should be taken to ensure that the points P0 , P1 both lie on the same side of the tangent P2 P3 , and similarly points P2 , P3 lie on the same side of the tangent P1 P0 . Setting a large value of ˛ violates this condition, resulting in a distorted Bezier curve. A second interpolation method that is suitable for vertex blending is Hermite interpolation. Here, the tangent directions are defined using vectors P0 Q0 and Q3 P3 (Fig. 4.4b), and the interpolating curve is given by H.t/ D.1 3t 2 C 2t 3 /P0 C .t 2t 2 C t 3 / ˛.P0 Q0 / C .t 2 C t 3 /˛.Q3 P3 / C .3t 2 2t 3 /P3

(4.4)

The coefficients of P0 , P3 are exactly same as that of Bezier interpolation. Since tangents are defined along the direction of the curve from P0 to P3 , Hermite interpolation does not have problems associated with large ˛ values. Hermite and other types of approximating splines are discussed in more detail in Chap. 7.

4.3 Skeleton and Skin Animating a three-dimensional character model (Fig. 4.1b) containing hundreds of vertices and polygons can be a challenging task. This task can be simplified to a great extent by grouping together a number of mesh vertices as forming body parts that move as a single unit, connected together by a set of joints. A human model may be modelled as a collection of body parts with joints at neck, shoulders, elbows, wrists, hips, knees, and ankles. The grouping of mesh primitives into body parts and the definition of joints depend on the complexity of the animation. In a simple walk sequence, for instance, the arms and legs could be considered as the only parts that move relative to the main body. For a more complex animation, one might require movement of the head, hands, fingers, facial muscle regions, and so on. Figure 4.5a shows how points in a mesh could be grouped into ten body parts: head (HEA), torso (TOR), left upper arm (LUA), left lower arm (LLA), right upper arm (RUA), right lower arm (RLA), left upper leg (LUL), lower left leg (LLL), right upper leg (RUL), and right lower leg (RLL). Every group can then have an abstract representation called a bone. The complete set of bones, along with their connectivity information, is called a skeleton (Fig. 4.5b). The notion of a skeleton consisting of a set of joint chains comprised of bones is central to articulated character animation. A skeleton can be easily animated; i.e., the transformations for the bones can be easily determined given the angles at each joint. The skeleton has the hierarchical structure similar to that of the model in Fig. 4.1a, the main difference being that in a skeleton, each component or bone is just an

58

4 Skeletal Animation

Fig. 4.5 (a) Vertices in a mesh model are grouped together into parts that move relative to each other. (b) A skeleton definition formed based on a vertex grouping

Fig. 4.6 Two simple ways of associating vertices with bones of a skeleton

abstract structure, not a graphics primitive. A bone essentially stores information about its position and orientation relative to its parent in the skeleton. Every bone is given a unique index as shown in Fig. 4.5b. Vertices belonging to a group are associated with a bone using the bone’s index. The part of a mesh represented by a bone is called its skin. In the example given in Fig. 4.5, the skin of bone “8” is the mesh segment that belongs to the set LUL. Two simple ways of associating groups of vertices with bones are shown in Fig. 4.6. In the first method, every entry in the vertex list is appended with a bone index. This method is suitable when vertices need to be associated with more than one bone (we will discuss this

4.4 Vertex Skinning

59

process later in Sect. 4.6). If several consecutive entries in the vertex list have the same bone index, then the second method is preferred where the minimum and maximum indices of a range of vertices are stored against a bone index.

4.4 Vertex Skinning In order to define the hierarchical nature of a skeleton, the parent–child relationship between every two connected bones must be shown. We could represent a bone using a point with arrow(s) pointing to its child node(s), as in Figs. 4.7a, c. Another common representation of a bone uses triangles. Fig. 4.7b shows the mesh model of a human arm, and the associated skeleton consisting of a set of bones. Each bone stores the index of its parent and the bone’s position relative to its parent. Using this information, a complete hierarchical structure can be built, as shown in Fig. 4.7c. There are two special nodes in this skeleton tree. The root node always represents the origin of the world coordinate system, and has an index 0. The base node is that bone in the skeleton which has root as its parent. The position and the orientation of the base define the pose of the skeleton in the world coordinate space. Bones are not physical structures present in a polygonal mesh, but are only animation tools or controlling mechanisms used to transform the mesh in a realistic manner. A bone also loosely represents the region of influence of a transformation.

4.4.1 The Bind Pose The hierarchical organization of bones in a skeleton allows the geometric transformation for each bone to be defined with respect to its parent. The transformations

a Parent

Child

c 2

4

1

3

5

Base

b 7

6 Base 8

9

10

Fig. 4.7 (a) A simple joint chain. (b) A skeletal structure for the arm, hand and fingers. (c) Modified version of the skeleton in Fig. 4.5b

60

4 Skeletal Animation

W V

θ

B′i Transformed bone matrix

Bi Initial bone matrix

Local coordinates frame of bone with index = i

Fig. 4.8 Transformation of a mesh vertex V using the transformations of its bone

that are associated with a bone are normally a joint angle rotation followed by a translation from its parent bone. The translations of the bones, each relative to its parent, together define the initial configuration of a skeleton. For this configuration, the joint angles are set to 0. The corresponding mesh is said to be in the bind pose (Fig. 4.5a). The placement of bones in the skeleton can be obtained by first computing the axis-aligned bounding boxes (see Box 3.1 of Chap. 3) of vertex groups (defined as in Fig. 4.6), and then determining the joint positions for each box. Fig. 4.5b shows an example. For now, we will assume that each vertex is attached to one and only one bone. We will consider a more general case of associating a vertex with two or more bones, in Sect. 4.6. In the following section, we will see how joint angle transformations applied to bones can be transferred to mesh vertices attached to them.

4.4.2 Mesh Vertex Transformation Consider a mesh vertex V attached to a bone i in bind pose (Fig. 4.8). In this configuration, each bone has an associated matrix Bi that defines the transformation from the bone’s local coordinate space to the skeleton’s coordinate space.

4.4 Vertex Skinning Fig. 4.9 An example showing transformations using three bones. (a) Bind pose and (b) transformed pose

61

a

y

V1

V3

V2

Bone-1

Bone-2

Bone-3

x z

b

W3

q3

W2 q2 y

W1 q1

x z

This transformation depends only on the translations of bones in the hierarchy relative to their parents. The process of obtaining this transformation matrix will be discussed below. For a given joint rotation by an angle , the transformed configuration of the bone in the skeleton’s coordinate space is represented by another bone matrix B0i . To get the transformed vertex W, we transfer the original point V from the coordinate space of the mesh (which is the same as the skeleton space) back to its bone’s local coordinate space, and then apply the joint angle transformation to return to the skeleton space. In other words, the vertex V is first transformed using the inverse of the matrix Bi , then by B0i . The first transformation gives the point Bi 1 V. Applying the matrix B0i to this point yields coordinates of the transformed point W. Thus W D .B0i Bi 1 /V

(4.5)

The above equation is fundamental to skeletal animation, as it describes how transformations applied to a bone i can be propagated to an attached mesh vertex V. The matrix Bi depends only on the initial configuration of the skeleton, and therefore the points Bi 1 V can be pre-computed and used for the entire animation sequence. As an example, we consider the model in Fig. 4.9, and show how it can be transformed using a skeleton comprising of three bones.

62

4 Skeletal Animation

Let d1 denote the translation vector used for moving Bone-1 from its local coordinate space to the skeleton space. Let d2 denote the vector by which Bone-2 is translated in Bone-1’s coordinate space. The vector d3 similarly represents the translation of Bone-3 in the coordinate space of Bone-2. Vertices V1 , V2 , V3 are attached to Bone-1, Bone-2, and Bone-3 respectively on the mesh in its bind pose (Fig. 4.9a). We seek to find the transformed coordinates of these vertices, when the skeleton is transformed using joint angles 1 , 2 , 3 respectively as shown in Fig. 4.9b. If we represent translation matrices by T, the initial bone matrices are given by B1 DT.d 1 / B2 DT.d 1 / T.d 2 / D T.d 1 C d 2 / B3 DT.d 1 /T.d 2 /T.d 3 / D T.d 1 C d 2 C d 3 /

(4.6)

When the bones are transformed using the joint angles, the bone matrices for the transformed configuration become B0 1 DT.d 1 / R.1 / B0 2 DT.d 1 / R.1 / T.d 2 / R.2 / B0 3 DT.d 1 / R.1 / T.d 2 / R.2 / T.d 3 / R.3 /

(4.7)

where R denotes a rotational transformation matrix. Now applying Eq. 4.5, we can write the expressions for the transformed vertex coordinates as W1 DT.d 1 / R.1 / T.d 1 / V1 W2 DT.d 1 / R.1 / T.d 2 / R.2 / T.d 1 d 2 / V2 W3 DT.d 1 / R.1 / T.d 2 / R.2 / T.d 3 / R.3 / T.d 1 d 2 d 3 / V3

(4.8)

So far we have assumed that each vertex is associated with only a single bone. Section 4.6 discusses a more general case.

4.5 Vertex Skinning Using Scene Graphs The vertex transformations (Eqs. 4.6, 4.7, 4.8) given in the previous section can be implemented using a scene graph for the skeleton. The scene graph is slightly different to the one we saw earlier in Chap. 3 (Fig. 3.3), in that each group node represents a bone with a matrix of the form M D TR defining the relative transformation of the bone with respect to its parent. Each bone has a child node representing the set of mesh vertices associated with that bone. In Fig. 4.10, Bone-1, Bone-2, and Bone-3 form a joint chain in a skeleton, and S2 denotes a

4.5 Vertex Skinning Using Scene Graphs

a

Root

b

Base node Bone-1

S1

63

Root

M1=T(d1)

Bone-2 S2

M1¢

Bone-1

M2=T(d2)

Bone-3

S1¢

M3=T(d3)

Bone-2 S2¢

M2¢

Bone-3

M3¢

S3¢

S3

Fig. 4.10 (a) Scene graph of a joint chain used for the pre-processing phase. (b) The updated mesh vertices in the animation phase

set of mesh vertices associated with Bone-2. The initial bone matrix for Bone-3 is B3 D M1 M2 M3 . The vectors Bi 1 V in Eq. 4.5 are obtained in a pre-processing phase, where each vertex set is transformed using the inverses of the matrices attached to nodes. As shown in Eq. 4.6, these matrices involve only translation components, and their inverses (as well as the product of inverses) can be easily computed. In the example given in Fig. 4.10a, a vertex V belonging to the set S3 would be transformed into 1 1 V 0 D M1 3 M2 M1 V D V d 1 d 2 d 3

(4.9)

As the tree is traversed from the root, matrices are combined by pre-multiplying the current product by the inverse of the matrix at the node, until a leaf node is reached. The vertices in a leaf node are transformed using the product of matrix inverses gathered up to that point. Thus the set S3 becomes a new set S3 0 after the transformation in Eq. 4.9. The transformed set of vertices replaces the original set for the animation phase (Fig. 4.10b). In the animation phase, matrices at scene graph nodes are updated using the joint angles of the bones. The updated matrices are represented by M0 in Fig. 4.10b. The scene graph is again traversed from the root; matrices are combined, this time using post-multiplication, and applied to the vertices at leaf nodes to get the transformed mesh vertices. The vertices in the set S3 0 would transform according to the following equation: W D M1 0 M2 0 M3 0 V 0

(4.10)

If the set of vertices attached to each bone can be specified as a range of indices (i, j) where i is the start index and j the end index of the set as in Fig. 4.6, then the structure of the scene graph can be simplified to a great extent as shown in Fig. 4.11. The vertex indices in the pre-processing phase point to the initial vertex list fVg

64

4 Skeletal Animation

Root

Root

Bone-1

M1 (i1 , j1)

M1′ (i1 , j1)

Bone-1

Bone-2

M2 (i2 , j2)

M2′ (i2 , j2)

Bone-2

Bone-3

M3 (i3 , j3)

M3′ (i3 , j3)

Bone-3

{V } {V ′ } {W } Vertex lists

Fig. 4.11 Simplified scene graph for a joint chain using a vertex index range for each bone

of the mesh. After the pre-processing phase, they point to the list of intermediate vertices fV 0 g that are used as inputs in the animation phase. The transformed list of vertices fWg is used for rendering the mesh after applying joint angle rotations to the bones (Fig. 4.11).

4.6 Transformation Blending If every vertex is attached to only a single bone, then transformations applied to the bones may cause mesh surfaces to interpenetrate at a joint (Fig. 4.12a, b). Figure 4.12b also shows how large flat surface patches can appear at a joint when two adjacent vertices move away from each other because of a rotational transformation. It is intuitive to transform vertices in the neighbourhood of a joint using a combination of bone matrices which influence that joint. If i and j are two bones that influence a joint, then a vertex V in the vicinity of the joint may be transformed using a weighted combination of the bone’s matrices Bi and Bj . The weights wi and wj are usually selected based on the relative distances of the vertex from the bones (Fig. 4.13). The final transformed point W (Fig. 4.12c) is obtained as ˚ W D wi B0i Bi 1 C wj B0j Bj 1 V

(4.11)

4.6 Transformation Blending

a

65

V

b

c

j

i

W

Mesh

Fig. 4.12 (a) A joint formed by two bones, and the attached mesh. (b) Interpenetration of mesh surfaces at a joint. (c) Mesh transformation using a combination of two bone matrices

i

(1.0, 0.0)

j

(0.9, 0.1)

(0.7, 0.3)

(0.5, 0.5)

(0.3, 0.7)

(0.1, 0.9)

(0.0, 1.0)

Fig. 4.13 Multiple weights associated with vertices for combining bone matrices

We also require the weights to satisfy the condition wi C wj D 1. A sample distribution of weights for mesh vertices of the joint in Fig. 4.12a is shown above (Fig. 4.13). In general, if n bones with indices 1, 2, : : : , n meet at a joint, the vertices surrounding the joint may be transformed using a matrix MD

n X i D1

wi B0 Bi 1

where;

n X

wi D 1:

(4.12)

i D1

The method outlined above is called transformation blending, and it usually produces smooth mesh deformations near joints. However when the angle of rotation of a bone is very large compared to its parent, the averaging scheme in Eq. 4.11 can produce two types of undesirable artefacts shown in Fig. 4.14. The first one is called a collapsing elbow effect, which appears when the angle between the axes of two adjacent bones becomes small. In this situation, vertex points on the inner edge of the mesh that are located near the joint move towards the centre. The second type of artefact is called the candy-wrapper effect, where one of the bones is twisted by 180ı about its axis. In this case, vertices with nearly equal weights get transformed to closely located points near the joint.

66

4 Skeletal Animation

b

a

Fig. 4.14 (a) Collapsing elbow effect. (b) Candy-wrapper effect

4.7 Keyframe Animation The animation of an articulated character model is usually done by specifying a set of keyframes that contain the information about the required joint transformation parameters at certain discrete points in time. Keyframes are generally predefined by an artist or an animator who can clearly specify the motion an object is required to produce. The joint angles for a character model at various instances in an animation sequence can also be obtained from motion capture systems. Here, the actions performed by a human actor are captured through the placement of markers near each joint of the body, and their recorded positions used to compute joint angles. A keyframe is essentially a time stamp of important transformation parameters and, optionally, other attributes such as colour, transparency etc. that are needed to render one frame of an animation sequence. As an example, consider a keyframe

a

b

Y

X

0

1

Neck

Body Centre

2

Right Shoulder

3

Left Shoulder

8

Right Hip

9

Left Hip

4

Right Elbow

5

Left Elbow

10

Right Knee

11

Left Knee

6

Right Wrist

7

Left Wrist

12

Right Ankle

13

Left Ankle

Z

Fig. 4.15 (a) Model of a stick figure and (b) the joint chains used for its animation. The hierarchical structure of links consists of five branches (chains), and 14 internal nodes. A leaf node is indicated by a blank square

4.7 Keyframe Animation

67

Fig. 4.16 The four primary keyframes used for generating a walk sequence of the stick figure. The graphs show the values of some of the joint angles, and a linear interpolation between the values is indicated by dotted lines

animation of the model of a stick figure shown in Fig. 4.15a. This model has five joint chains, and a total of 14 joints (Fig. 4.15b). A single configuration or “pose” of the model is therefore given by 14 joint angles 0 , 1 , : : : , 13 , and the position (x0 , y0 , z0 ) of the root joint. A joint rotation that moves a link forward (towards C z) is considered as positive. For example, the elbow joints are constrained to rotate the arm only forward, by assigning only positive values for 4 and 5 . Similarly the knee joint angles ( 10 , 11 ) are always assigned a negative value. An alternative definition for these joint angles can be obtained by viewing them as rotations about the x-axis. In this case, the angles at shoulders and elbows will have negative values, and the angles at the knees will have positive values. For a simple walk sequence for the stick figure, four key-frames are defined as shown in Fig. 4.16. These are the primary postures from which the intermediate motion can be generated by linear interpolation. In our example, movements of the

68

4 Skeletal Animation

a 10

-10

Step interpolation

b

10

-10

Linear interpolation

c

10

-10

Spline interpolation

Fig. 4.17 Commonly used interpolation methods in keyframe animation

neck, shoulder and wrists are neglected, and hence values of only ten joint angles are specified for each keyframe. More complex and realistic movements such as running, jumping or performing somersaults can be produced by creating a larger number of keyframes using motion capture systems. The “in-between” frames of an animation sequence are generated by interpolating keyframe values using either step, linear or spline functions. A step function uses the previous keyframe values for all subsequent frames until another keyframe is encountered (Fig. 4.17a). A linear interpolation produces points along line segments connecting two consecutive keyframes (Fig. 4.17b). If k1 denotes a parameter in a keyframe at time t1 , and k2 denotes the value of the same parameter in the next keyframe at time t2 , the value k for an in-between frame at time t is given by kD

t2 t t2 t1

k1 C

t t1 t2 t1

k2

(4.13)

0 1:

(4.14)

or equivalently, k D .1 /k1 C k2 ; where D

t t1 ; t2 t1

(4.15)

For smoother motion interpolation, keyframe values are connected using piecewise cubic splines (Fig. 4.17c). Catmull-Rom splines are commonly used for this purpose, as they have properties of both C0 and C1 continuity between consecutive

4.8 Sample Implementation of Vertex Skinning

69

spline segments. Please refer to Chap. 7 (Sects. 7.2 and 7.5) for more information on Catmull-Rom and other types of splines that are useful for generating approximating curves and surfaces.

4.8 Sample Implementation of Vertex Skinning In Sect. 3.5 we discussed the implementation of a scene graph class. For vertex skinning, we will use a highly simplified model where the information attached to each node is appended with a vertex index range given by the first and the last indices of the range. In this model, there is no need for an object node, and the vertices are processed at group nodes only. Listing 4.1 gives the class definition for a skeleton node. A documentation of the methods in this class can be found in Appendix C. Just like a scene graph node, a skeleton node also stores transformation parameters and a list of pointers to its child nodes.

4.8.1 Skeleton Node The primary function of the SkeletonNode is to provide a convenient framework for representing the bone hierarchy and also to transform the vertex list of a mesh model using joint angles specified for each bone. The two functions preprocessPhase() and animationPhase() both initiate a recursive traversal of the tree to transform entries in the vertex lists as shown in Fig. 4.11. It is useful to have a Skeleton class with functions to load a skeleton definition and to define joint angles for bones during animation (Listing 4.2). These two functions provide the main interface between the classes and the user application. A Skeleton object represents the whole skeleton of a mesh model consisting of several bones (skeleton nodes). The contents of the skeleton definition file are organized as shown in Fig. 4.18. The loadSkeleton() function reads in the parameters and builds the hierarchical structure. The reference to the root node is available to the application via the function getRoot().

4.8.2 Skinned Mesh Node The SkinnedMesh class encapsulates data and related functions for loading a mesh file consisting of vertex and polygon lists, attaching a skeleton, and transforming the vertices using the joint angles associated with the bones of the skeleton (Listing 4.3). As shown in Fig. 4.11, the SkinnedMesh class uses three vertex lists in the form of vectors to store the initial vertices of the mesh in bind pose, the intermediate set of vertices after the pre-processing phase, and the final set of vertices after

70

4 Skeletal Animation Listing 4.1 Class definition for a single node of a skeleton

applying joint angle transformations. The mesh definition file has a simple structure consisting of the list of vertices and polygons. Polygons are specified using vertex indices (three indices for triangles and four for quads). The vertex index starts from 1. Figure 4.19 gives the mesh definition for a rectangular prism. The framework described above also uses the Point3 and Matrix classes for various vertex and transformation related functions (see Appendix A). This book’s

4.8 Sample Implementation of Vertex Skinning

71

Listing 4.2 Class definition for a skeleton

Fig. 4.18 Sample skeleton definition file

companion website contains the header and implementation files of all the above classes. An example of a simple application using the skeleton animation framework is shown in Listing 4.4. At the initialization stage, both mesh and skeleton objects are created, corresponding data loaded from input files, the skeleton is attached to the mesh, and the preprocess() function is called on the mesh object. This function in turn passes the vertex data to the root node of the skeleton via the preprocessPhase()function and gets back the intermediate vertices. The display() function performs the animation of the mesh by defining joint angles for the bones. In the example, the function call skeleton->rotate(3,30,0,-75) is used to rotate the bone with index 3 by 30ı about the x-axis and 75ı about the z-axis. The sequence of rotations is

72

4 Skeletal Animation Listing 4.3 Class definition for a mesh

pre-defined. The function call mesh->render() is used inside the display loop to render the mesh with the transformed vertex coordinates.

4.9 Summary This chapter addressed the problem of animating articulated character models. Character models are divided into two main categories: those constructed using individual component objects, and those modelled as a single mesh surface. The first category of objects requires blending of surfaces at the joints to avoid interpenetration of component objects and the appearance of gaps during animation.

4.9 Summary

73

Fig. 4.19 A sample mesh definition file Listing 4.4 Example of an application using the vertex skinning algorithm

It was shown that Hermite polynomials and cubic Bezier polynomials could be effectively used for vertex blending. This chapter also presented the vertex skinning algorithm which is a well-known method used in skeletal animation. Various aspects of vertex skinning including

74

4 Skeletal Animation

the transformation of mesh vertices using skeletons, application of scene graphs in vertex skinning, and transformations using a combination of bone matrices have been discussed in detail. The process of keyframe interpolation has been outlined. This chapter also demonstrated the implementation of the vertex skinning algorithm. The next chapter introduces the quaternion algebra and transformations that are used for interpolating between orientations in three-dimensional space. Quaternions have a very important role in animation sequences where generic rotational transformations are applied to objects.

4.10 Supplementary Material for Chap. 4 The folder Chapter4/Code on the companion website contains code examples demonstrating the application of concepts introduced in this chapter. A brief description of these programs is given below. 1. SkeletonNode.cpp

This class implements the basic functionalities of a scene graph for skeleton animation as detailed in Sect. 4.8. The class documentation can be found in Appendix C. 2. SkinnedMesh.cpp

This class supports several functions for loading and rendering a skinned mesh file. A brief description of the class can be found in Sect. 4.8.2, and the class documentation in Appendix C.

4.10 Supplementary Material for Chap. 4

75

3. VertexBlending.cpp

This program generates a blending surface between two cylinders using Hermite interpolation. Clicking the left mouse button starts the rotation of one of the cylinders. Use up or down arrow keys to increase or reduce the weight ˛ of the tangent vectors. Press left or right arrow keys to change the view direction. 4. TwoBoneTransform.cpp

The program demonstrates the collapsing elbow and candy wrapper effects seen in transformations using a combination of two bone matrices. Use left and right arrow keys to increase or decrease the bending angle (rotation about the z-axis). Use up and down arrow keys to decrease or increase the twist angle (rotation about the x-axis). The spread of the weights can be increased by pressing the ‘s’ key, and decreased by pressing the ‘a’ key. 5. HumanModel.cpp

This program uses the vertex skinning method to transform a mesh based on transformations applied to a skeleton. It requires two input files, “HumanModel.txt” (mesh definition) and “Skeleton.txt” (skeleton definition). The bone

76

4 Skeletal Animation

indices are defined as given in Fig. 4.5. The bone transformations are defined inside the display() function of the program. Use left and right arrow keys to change the view direction.

4.11 Bibliographical Notes Both vertex blending and vertex skinning are often used synonymously in computer graphics literature. In this book, vertex blending refers to an interpolation method between polyhedral surfaces, while vertex skinning refers to a completely different method of animating a mesh using a skeleton. The process of constructing blending surfaces between polyhedral objects is often referred to as polyhedral vertex blending. Such methods were originally introduced for Computer Aided Design (CAD) applications. Bajaj and Ihm (1992) gives the fundamental concepts for designing blending surfaces with Hermite polynomials. A description of parametric cubic curves and surfaces generated using Hermite polynomials can be found in Foley (1994, 1996), and Angel (2008). Cubic interpolation methods using Hermite curves are discussed in Eberly (2007) and Moller et al. (2008). Skeleton animation is an important technique in game programming and character animation. Books such as Astle (2006), Moller et al. (2008) and Erleben (2005) provide a description of skeleton based mesh transformation methods. Eberly (2007) gives an outline of the vertex skinning method. The implementation aspects of vertex skinning are presented in Lander (1998) and Kavan (2003).

References Angel, E. (2008). Interactive computer graphics: A top-down approach using OpenGL (5th ed.). Boston/London: Pearson Addison-Wesley. Astle, D. (2006). More OpenGL game programming. Boston: Thomson/Course Technology. Bajaj, C. L., & Ihm, I. (1992). Algebraic surface design with Hermite interpolation. ACM Transactions on Graphics, 11(1), 61–91. Eberly, D. H. (2007). 3D game engine design: A practical approach to real-time computer graphics (2nd ed.). Amsterdam/London: Morgan Kaufmann. Erleben, K. (2005). Physics-based animation (1st ed.). Hingham: Charles River Media. Foley, J. D. (1994). Introduction to computer graphics (Abridged and modified edn.). Reading/ Wokingham: Addison-Wesley. Foley, J. D. (1996). Computer graphics: Principles and practice (2nd ed.). Reading/Wokingham: Addison-Wesley. Kavan, L. (2003). Real-time skin deformation with bones blending. International conference in central Europe on computer graphics, visualization and computer vision. Plzen, Czech Republic. Lander, J. (1998). Skin them bones. Game Developer, 5, 11–16. Moller, T., Haines, E., & Hoffman, N. (2008). Real-time rendering (3rd ed.). Wellesley: A.K. Peters.

Chapter 5

Quaternions

Overview In computer graphics applications, quaternions are used to represent threedimensional rotations. They provide some key advantages over the traditional way of defining generic rotational transformations using Euler angles. Quaternions are also extremely useful for interpolating between two orientations in threedimensional space. Keyframe animations requiring orientation interpolation therefore find a very convenient mathematical tool in quaternions. This chapter gives an overview of the algebra of quaternions, the geometrical interpretation of quaternion transformations, and quaternion based linear and spherical interpolation functions. A comparison of rotation interpolation methods using Euler angles, angle-axis representations, and quaternions is presented. The extension of quaternions to eight-dimensional dual quaternions and their usefulness in representing general rigid-body transformations are also discussed.

5.1 Review of Complex Numbers Quaternions are hyper-complex numbers of rank 4, and therefore it is useful to review some of the basic concepts related to complex number algebra to gain a better insight into quaternion operations. Even though a complex number z is p commonly represented in the form a C i b where i D 1, and a, b are respectively the real and imaginary parts of z, we will use the two-tuple notation (a, b) for z. With this notation, we can write 1 D (1, 0), and i D (0, 1). These two-dimensional vectors (1, 0) and (0, 1) form an orthogonal basis for the complex space, where any number z D (a, b) can be expressed as their linear combination a (1, 0) C b (0, 1). The operations of addition, subtraction and multiplication in the field of complex numbers are defined as follows: .a1 ; b1 / ˙ .a2 ; b2 / D .a1 ˙ a2 ; b1 ˙ b2 / R. Mukundan, Advanced Methods in Computer Graphics: With examples in OpenGL, DOI 10.1007/978-1-4471-2340-8 5, © Springer-Verlag London Limited 2012

(5.1) 77

78

5 Quaternions

Fig. 5.1 Multiplication by a unit complex number has the effect of rotation of vectors and points about the origin on a two-dimensional plane

y (x cosd -y sind, x sind+y cosd)

p¢ (x,y) d

p x

.a1 ; b1 /.a2 ; b2 / D .a1 a2 b1 b2 ; a1 b2 C a2 b1 /

(5.2)

c.a; b/ D .ca; cb/;

(5.3)

where c is a real number. The multiplication rule given in Eq. 5.2 establishes the fact that i2 D (0,1) (0,1) D (1, 0). The complex conjugate of z D (a, b) is given by z* D (a, b). The magnitude of z is a positive real number defined as jzj D

p a2 C b 2

(5.4)

Using the multiplication rule, we find that zz D jzj2 D a2 C b 2

(5.5)

If a complex number z has a unit magnitude, then zz* D 1. This implies that for a unit complex number, z* is the multiplicative inverse of z. All unit complex numbers can be expressed in the general form z D .cosı; sinı/

(5.6)

Consider any vector (or point) p D (x, y) in a two-dimensional coordinate system. If we treat p as a complex number, and multiply it by the unit complex number z given above, the product zp can be evaluated using Eq. 5.2 as follows: p0 D .cosı; sinı/.x; y/ D .xcosı ysinı; xsinı C ycosı/

(5.7)

The transformed vector (or point) p0 has the same magnitude as p, and can be obtained by rotating p about the origin by an angle ı (Fig. 5.1). The unit complex vector therefore represents a rotation in two-dimensional space. The geometrical interpretation of unit complex numbers as rotation operators forms the basis for the framework for an extended set of hyper-complex numbers

5.2 Quaternion Algebra

79

called quaternions. We will see shortly that unit quaternions represent threedimensional rotations. In the following section, we introduce the algebra of quaternion numbers.

5.2 Quaternion Algebra We have seen above that the field of complex numbers have 1 D (1, 0), i D (0, 1) as the orthogonal basis. The quaternion set has an extended orthogonal basis consisting of four elements 1 D (1, 0, 0, 0), i D (0, 1, 0, 0), j D (0, 0, 1, 0), k D (0, 0, 0, 1). Thus a quaternion Q D (q0 , q1 , q2 , q3 ) has an equivalent representation q0 C q1 i C q2 j C q3 k, where the quaternion components qi are all real values. The term q0 is called the scalar part of Q, and the 3-tuple (q1 , q2 , q3 ) the vector part. The operations of addition, subtraction and scalar multiplication are defined as follows: .p0 ; p1 ; p2 ; p3 / ˙ .q0 ; q1 ; q2 ; q3 / D .p0 ˙ q0 ; p1 ˙ q1 ; p2 ˙ q2 ; p3 ˙ q3 / (5.8) c.q0 ; q1 ; q2 ; q3 / D .cq0 ; cq1 ; cq2 ; cq3 /;

(5.9)

where c is any real number. Analogous to Eq. 5.2, the quaternion product is given by .p0 ; p1 ; p2 ; p3 /.q0 ; q1 ; q2 ; q3 / D .p0 q0 p1 q1 p2 q2 p3 q3 ; p0 q1 C p1 q0 C p2 q3 p3 q2 ; p0 q2 p1 q3 C p2 q0 C p3 q1 ; p0 q3 C p1 q2 p2 q1 C p3 q0 /

(5.10)

From the above definition of a quaternion product, it is obvious that quaternion multiplication is not commutative. That is, for any two quaternions P D (p0 , p1 , p2 , p3 ), Q D (q0 , q1 , q2 , q3 ), the product PQ need not necessarily be the same as QP. If we denote the vector part of P by v D (p1 , p2 , p3 ) and the vector part of Q by w D (q1 , q2 , q3 ), then Eq. 5.10 becomes .p0 ; v/.q0 ; w/ D .p0 q0 v w; p0 w C q0 v C v w/

(5.11)

where v•w denotes the dot product and vw the cross product of the two vectors. The right-hand side of Eq. 5.10 when treated as a column vector, can be conveniently expressed as a product of a matrix of elements of P and a vector containing elements of Q as given below. 2

p0 6 p1 PQ D 6 4 p2 p3

p1 p0 p3 p2

p2 p3 p0 p1

32 3 p3 q0 6 q1 7 p2 7 76 7; p1 5 4 q2 5 p0 q3

(5.12)

80

5 Quaternions

or, equivalently as 2

32 3 q0 q1 q2 q3 p0 6 q1 q0 q3 q2 7 6 p1 7 76 7 PQ D 6 4 q2 q3 q0 q1 5 4 p2 5 q3 q2 q1 q0 p3

(5.13)

From Eq. 5.10, we can derive the following properties satisfied by the quaternion basis: i 2 D j 2 D k 2 D ij k D 1 ij D j i D k j k D kj D i ki D i k D j

(5.14)

Quaternions also form a commutative group under addition, where (0,0,0,0) is the identity element. Quaternion multiplication is associative, and distributes over addition. If P, Q, R are any three quaternions, .PQ/R D P .QR/ .P C Q/R D PR C QR P .Q C R/ D PQ C PR

(5.15)

The conjugate Q* of the quaternion Q D (q0 , q1 , q2 , q3 ) is defined as Q D .q0 ; q1 ; q2 ; q3 /

(5.16)

Thus, if Q D (q0 , w), then Q* D (q0 , w). Also, Q C Q* D 2q0 . The magnitude (also called the length, or norm) of Q denoted by jQj, is jQj D

q q02 C q12 C q22 C q32

(5.17)

By taking the magnitude of the quaternion product in Eq. 5.10 we get jPQj D jP jjQj

(5.18)

Using Eq. 5.11, it is easy to find that QQ D Q Q D jQj2 :

(5.19)

5.3 Quaternion Transformation

81

By dividing the above equation by jQj2 , we get the equation for the quaternion inverse. If we denote the quaternion inverse of Q by Q1 , then Q1 D

Q jQj2

(5.20)

A quaternion Q can be normalized to a unit quaternion by dividing each of its components by the length jQj given in Eq. 5.17. A unit quaternion satisfies the following equations: jQj D 1: q0 2 C q1 2 C q2 2 C q3 2 D 1: Q1 D Q

(5.21)

If the real part q0 of a quaternion is zero, it represents a vector (q1 , q2 , q3 ) in three-dimensional space. Such a quaternion that has the form (0, q1 , q2 , q3 ) D (0, q) is called a pure quaternion. Similarly, quaternions of the type (a, 0, 0, 0) with the vector component zero are called real quaternions. The algebra of real quaternions is the same as that of real numbers. Similarly, quaternions of the type (a, b, 0, 0) behave exactly like complex numbers (a, b).

5.3 Quaternion Transformation A special type of quaternion product in the form QPQ* plays an important role in three-dimensional transformations. We have just seen that a vector p in the threedimensional space corresponds to a pure quaternion P D (0, p). An interesting fact that leads to the notion of a quaternion transformation is that given any quaternion Q and a pure quaternion P, the product P0 D QPQ* is also a pure quaternion. Thus QPQ* can be viewed as the transformation of a pure quaternion P D (0, p1 , p2 , p3 ) using another quaternion Q. We can derive the matrix form of this transformation by using Eq. 5.13 for obtaining the matrix expression for PQ* and then using Eq. 5.12 for getting the final product Q(PQ*). 32 q0 q1 q2 q3 q0 q1 q2 q3 7 6 6 q q q q q 1 0 3 2 76 1 q0 q3 q2 QPQ D 6 4 q2 q3 q0 q1 5 4 q2 q3 q0 q1 q3 q2 q1 q0 q3 q2 q1 q0 2

32

3 p0 7 6 p1 7 76 7 5 4 p2 5 p3

(5.22)

82

5 Quaternions

The following matrix equation immediately follows by multiplying the two matrices together, and setting p0 D 0: 2

0

3

2

1

0

0

0

2 .q0 q1 C q2 q3 /

q02 q12 q22 C q32

6 0 7 6 6 p 1 7 6 0 q02 C q12 q22 q32 2 .q0 q3 C q1 q2 / 2 .q0 q2 C q1 q3 / 6 7 6 6 0 7D6 4 p 2 5 4 0 2 .q0 q3 C q1 q2 / q02 q12 C q22 q32 2 .q0 q1 C q2 q3 / p03

0 2 .q0 q2 C q1 q3 /

32

0

3

76 7 7 6 p1 7 76 7 76 7 5 4 p2 5 p3 (5.23)

This equation defines the quaternion transformation of a three-dimensional point (or vector) p D (p1 , p2 , p3 ) to another three-dimensional point (or vector) p0 D (p1 0 , p2 0 , p3 0 ). An alternative form of the equation can be derived as follows: QPQ D .q0 ; w/.0; p/.q0 ; w/;

(5.24)

where w D (q1 , q2 , q3 ) . Using Eq. 5.11 to expand the product term, we get QPQ D .0; q0 2 p C w.p w/ C 2q0 .w p/ C w .w p//

(5.25)

The above equation proves that the transformation of P is also a pure quaternion. We can therefore write p0 D q0 2 p C w.p w/ C 2q0 .w p/ C w .w p/

(5.26)

Further simplification of the right-hand side using vector algebra gives p0 D .q0 2 w2 / p C 2w.p w/ C 2q0 .w p/

(5.27)

where, w2 D jwj2 D q1 2 C q2 2 C q3 2 . It should be noted that QPQ* generally is not a scale-preserving transformation because jP 0 j D j Qj2 jP j

(5.28)

If we impose the constraint that Q is a unit quaternion (i.e., jQj D 1), we get a scale-invariant (or length-preserving) transform. With this additional criterion, we can also write the inverse quaternion transform in a concise form as P D QP 0 Q

(5.29)

We also note that when P is the zero-quaternion (0, 0, 0, 0), so is P0 . Therefore the origin is a fixed point of the transformation. A length-preserving transformation with a fixed point is a rotation. In the following sections we will attempt to find a

5.4 Generalized Rotations

83

geometric interpretation of the quaternion transformation as a pure rotation in threedimensional space, and express the components of a unit quaternion in terms of the angle and the axis of rotation.

5.4 Generalized Rotations Before we further analyze the transform properties of quaternions, it would be worthwhile to review some of the key concepts relating to general three-dimensional rotations. Any composite transformation that preserves length, angle and area is called a rigid-body transformation. If a rigid body transformation has also a fixed point (pivot), then it is a rotation. A rotation can be measured in terms of the angular deviation of an orthogonal right-handed system fixed on the rotating body, with the origin of the system at the fixed point of rotation. In Fig. 5.2a, Ox, Oy, Oz are the axes of an orthogonal triad before rotation, and Oxt , Oyt , Ozt denote the transformed axes directions after a rotation about O. The coordinate reference frame is inertially fixed and is represented by X, Y, Z axes. A general rigid body transformation of an object without a fixed point can be treated as a rotation followed by a translation. Such a transformation can be equivalently performed by first carrying out a rotation that aligns the axes parallel to the final directions, followed by a translation that moves the fixed point O to its final position Ot (Fig. 5.2b). While any translation can be unambiguously represented by a three component vector, a general rotation may be specified in several ways. In the following, we consider the Euler angle and angle-axis representations of threedimensional rotations.

a

b

Y

zt Ot

y

xt yt

xt zt Y O

X

z

z

O

x yt Z

y X

x Z

Fig. 5.2 (a) A generalized rotation with a fixed point O that transforms the directions of bodyfixed axes from O(x, y, z) to O(xt , yt , zt ). (b) A general transformation without a fixed point

84

5 Quaternions

5.4.1 Euler Angles The Euler’s theorem on rotations states that any general rotation can be performed using a sequence of elementary rotations about the coordinate axes passing through the fixed point. The theorem further states that if no two successive rotations is about the same axis, then the maximum number of rotations needed to achieve the transformation is three. Thus any rotational transformation can be represented by a sequence of three rotations about mutually independent axes. These angles are called Euler angles. Before defining an Euler angle representation, we need to fix the sequence in which the rotations are performed. If we denote rotations about the X-axis by , rotations about Y by , and rotations about Z by , a set of Euler angles can be defined using any of the following 12 sequences: 1 1

2 2

1 2 1 2

1 2 1 2

The Euler angle sequence f g represents a rotation about X followed by a second rotation about Y, followed by a third rotation about the Z axis. The sequence f 1 2 g gives another Euler angle representation in terms of a rotation about the Y axis, followed by a second rotation about the X axis, and then a third rotation again about the Y axis. The six sequences where each axis is used exactly once are called proper Euler angles. The transformation matrix for the f g sequence is obtained by concatenating the transformation matrices as shown below. 2 6 6 6 4

x0 y0 z0 1

3

2

7 6 7 6 7 D6 5 4 2

cos sin sin cos 0 0 0 0

cos ' cos sin 6 6 cos ' sin sin D6 4 sin ' 0

0 0 1 0

0 0 0 1

32 76 76 76 54

cos ' 0 sin ' 0

sin ' cos cos sin ' sin C cos sin cos ' 0

3 x 76 7 76 76 y 7 76 76 7 76 54 z 5 54 1 32 sin cos sin ' cos C sin sin 0 x 76 cos cos sin ' sin sin cos 0 7 6 y 76 cos cos ' 0 54 z 0 1 1

0 sin ' 0 1 0 0 0 cos ' 0 0 0 1

32

1 0 0 cos 0 sin 0 0

0 sin cos 0

0 0 0 1

32

3 7 7 7 5

(5.30) The above equation can be interpreted as the transformation of any point (x, y, z) to (x0 , y0 , z0 ) in a fixed coordinate frame. This interpretation does not use any information pertaining to body-fixed axes. On the other hand, if we assume that x, y, z represent the body-fixed axes which initially coincide with the coordinate reference axes X, Y, Z, respectively, Eq. 5.30 can be viewed as the transformation of a point from the moving body frame to the fixed coordinate reference frame. The Euler angle representation described above (and shown in Fig. 5.3) used rotations

5.4 Generalized Rotations

85

Fig. 5.3 An extrinsic composition of Euler angle rotations performed using the sequence f , , g

Fig. 5.4 An intrinsic composition of Euler angle rotations performed using the sequence f , , g

that are performed about the fixed principal axes directions X, Y, Z of the reference frame. Such a transformation is called an extrinsic composition of rotations. An intrinsic composition, on the other hand, uses rotations about body-fixed axes whose directions change in the reference frame after every rotation. For example, an aircraft orientation is defined in this manner. In Fig. 5.4, the yaw rotation is performed about the x-axis, the roll rotation about the transformed body y-axis, and the pitch rotation about the transformed body z-axis. For this sequence of intrinsic composition of rotations, the transformation from body frame to the coordinate reference frame is given by 3 2 1 0 0 X 6 Y 7 6 0 cos sin 6 7D6 4 Z 5 4 0 sin cos 0 0 0 1 2 cos sin 6 sin cos 6 4 0 0 0 0 2

32 cos 0 6 0 07 76 0 5 4 sin 0 1 32 3 0 0 x 6y7 0 07 76 7 1 0 54 z 5 0 1 1

3 0 sin 0 1 0 07 7 0 cos 0 5 0 0 1

(5.31)

86

5 Quaternions

Fig. 5.5 Two different Euler angle interpolation sequences generated for the same initial and target orientations Fig. 5.6 Transformation of a vector under a general rotation about the origin in three-dimensional space

u = (l, m,n)

s

r

P

d P¢ p

p¢ a t

O

A three-dimensional orientation can be represented in different ways using different Euler angle sequences. Even if we keep the sequence fixed, certain orientations can have more than one set of Euler angles. For instance, using the same sequence f g, both f45, 80, 0g and f135, 100, 180g represent the same transformation. This can be verified by evaluating the product matrix in Eq. 5.30 for the two sets of angles. The non-uniqueness of the Euler angle representation also means that you may not get a unique interpolation path between two orientations (Fig. 5.5).

5.4.2 Angle-Axis Transformation The Euler’s theorem concerning three-dimensional rotations states that any number of rotational transformations with a single fixed point applied to an object can be replaced by a single rotation of the object about an axis passing through the fixed point. The axis is often called the equivalent axis of rotation. Any orientation of an object with the origin as a fixed point can therefore be specified using an angle of rotation ı and an axis of rotation given by a unit vector u D (l, m, n). In the following discussion, we assume that the axis of rotation passes through the origin. Figure 5.6 depicts the rotational transformation applied to a vector p (or a point P). If we denote the projected lengths of the vector p along directions of u (axis of rotation) and s (perpendicular to axis of rotation) by a and r respectively, we can

5.4 Generalized Rotations

87

write p D au C rs, where a D p•u. During any rotation of the vector p about the axis u, both these projected distances a and r remain constant. If t denotes the vector orthogonal to both u and s, the transformed vector direction p0 can be written as p0 D au C .r cos ı/s C .r sin ı/t D au C .p au/cos ı C .u p/sin ı D p cos ı C .1 cos ı/.p u/u C .u p/sin ı

(5.32)

The above equation is the well-known Rodrigues’ rotation formula. The matrix version of the Rodrigues’ formula can be derived by defining a 3 3 skewsymmetric matrix UX as 2 3 0 n m Ux D 4 n 0 l 5 ; (5.33) m l 0 and replacing u, p, p0 by the corresponding column vectors: 2

3 l U D 4m5; n

2 3 x p D 4y 5; z

2

3 x0 p0 D 4 y 0 5 : z0

(5.34)

With the above notations, the vector cross-product up has an equivalent matrix representation (UX ) p. It can also be easily verified that the term (p•u) u in Eq. 5.32 is equivalent to the matrix (UUT ) p. Thus we get (5.35) p0 D I cos ı C .1 cos ı/UUT C UX sin ı p Noting that U2X D UUT C I

(5.36)

Equation 5.35 can be written in an alternate form as below. p0 D I C .1 cos ı/U2X C UX sin ı p

(5.37)

Equation 5.35 can also be written in the expanded matrix form as follows for defining the rotational transformation of a point P expressed in homogeneous coordinates: 2

x0

3

2

6 07 6 6y 7 6 6 7D6 6 07 6 4z 5 4 1

32 3 x 7 6 7 6y 7 lm.1 cos ı/ C n sin ı m2 .1 cos ı/ C cos ı mn.1 cos ı/ l sin ı 0 7 76 7 7 6 7 nl.1 cos ı/ m sin ı mn.1 cos ı/ C l sin ı n2 .1 cos ı/ C cos ı 0 5 4 z 5 0 0 0 1 1 l 2 .1 cos ı/ C cos ı

lm.1 cos ı/ n sin ı nl.1 cos ı/ C m sin ı 0

(5.38)

88

5 Quaternions

Let us consider the problem of computing the equivalent angle and axis of rotation from a transformation matrix. Given a general 44 rotation matrix in the form 2

m00 m01 m02 6 m10 m11 m12 6 4 m20 m21 m22 0 0 0

3 0 07 7 05 1

(5.39)

we get the following equations using the matrix elements from Eq. 5.38: m00 C m11 C m22 D 1 C 2cos ı m21 m12 D 2l sin ı m02 m20 D 2m sinı m10 m01 D 2nsinı

(5.40)

From the above equations, we can derive the expressions for angle and axis of rotation as follows: 0q 1 2 2 2 .m m / C .m m / C .m m / 21 12 02 20 10 01 C B ı D tan1 @ A m00 C m11 C m22 1 m21 m12 2 sin ı m02 m20 mD 2 sin ı m10 m01 nD 2 sin ı lD

(5.41)

In the next section, we will establish the equivalence between an angle-axis transformation and a unit quaternion transformation of the form QPQ* where P is a pure quaternion (0, p).

5.5 Quaternion Rotations We will now try to represent the rotational transformation in Fig. 5.6 by a unit quaternion Q D (q0 ,w), where the vector component w of the quaternion is along the axis of rotation. Therefore we have w D ku; for some constant k:

(5.42)

5.5 Quaternion Rotations

89

We saw earlier that a vector p can be transformed into another vector p0 using a unit quaternion Q and the result of this transformation is given by Eq. 5.27. In the previous section, we considered an angle-axis transformation of a vector p given by Eq. 5.32. We find a striking similarity between the two equations, which suggests that the quaternion transformation in Eq. 5.27 is indeed an angle-axis transformation. Equating the corresponding terms in both the equations, we find that q0 2 w2 D cos ı 2q0 k D sin ı 1 cos ı D 2k 2

(5.43)

From the above equations, we can see that k D sin(ı/2), and q0 D cos(ı/2). Therefore the unit quaternion that represents the rotation in Fig. 5.6 is given by ı ı ı ı Q D cos ; l sin ; m sin ; n sin 2 2 2 2

(5.44)

This result is fundamental to the theory of generalized rotations, as it provides a direct mechanism for converting angle-axis representations of three-dimensional rotations into unit quaternions. From this equation, we can also derive the relationship between the components of any unit quaternion Q D (q0 , q1 , q2 , q3 ) and the parameters of rotation it represents. The angle of rotation is given by 0q B ı D 2tan1 @

1 q12 C q22 C q32 C A q0

(5.45)

and the unit vector along the axis of rotation (l, m, n) can be obtained as q1 lD q q12 C q22 C q32 q2 mD q q12 C q22 C q32 q3 nD q q12 C q22 C q32

(5.46)

Replacing ı/2 with ı in Eq. 5.44, we can summarize our discussion above as follows: Any unit quaternion Q can be expressed in the form Q D (cosı, u sinı), and it represents a rotation by an angle 2ı about a unit vector u passing through the origin.

90

5 Quaternions

5.5.1 Quaternion Transformation Matrix From the above discussion, we can conclude that if Q is a unit quaternion, then Eq. 5.27 gives a rotational transformation of a vector p D (x, y, z, 0). This transformation equation could also be written in the conventional matrix form as shown below: 2 03 2 32 3 1 2q22 2q32 2q1 q2 2q0 q3 2q1 q3 C 2q0 q2 0 x x 6 07 6 76 7 2 2 6 y 7 6 2q1 q2 C 2q0 q3 1 2q1 2q3 2q2 q3 2q0 q1 0 7 6 y 7 6 7D6 76 7 6 z0 7 6 2q q 2q q 2q q C 2q q 1 2q 2 2q 2 0 7 6 z 7 (5.47) 0 2 2 3 0 1 4 5 4 1 3 54 5 1 2 0 0 0 0 0 1 The same transformation matrix can be applied to transform a point P D (x, y, z, 1) to another point P0 D (x0 , y0 , z0 , 1) using the quaternion Q. The quaternion transformation matrix in Eq. 5.47 is orthogonal, meaning that its inverse is the same as its transpose. The matrix also has some very useful properties. If we equate this matrix to a general 4 4 matrix given in Eq. 5.39, we can find that the following relationships hold among the matrix elements: m00 C m11 C m22 C 1 D 4q0 2 m21 m12 D 4q0 q1 m02 m20 D 4q0 q2 m10 m01 D 4q0 q3

(5.48)

The above equations are useful for extracting the quaternion elements from a given 4 4 rotational transformation matrix: q0 D q1 D q2 D q3 D

p 1 C m00 C m11 C m22 2 m21 m12 4q0 m02 m20 4q0 m10 m01 4q0

(5.49)

We will choose only the positive value of the square-root for computing q0 . A negative value for q0 will change the sign of all remaining components and yield the quaternion Q in place of Q. Shortly (Eq. 5.57) we will see that both Q and Q represent the same rotation, and therefore we can safely impose the constraint

5.5 Quaternion Rotations

91

that the sign of q0 is positive, and compute the remaining components from it. Note also that the above equations are valid only when q0 ¤ 0. If q0 D 0, then the angle of rotation ı D ˙180ı , and the matrix in Eq. 5.47 becomes a symmetric matrix. For this special case, the remaining quaternion elements can be derived as follows: p 1 C m00 m11 m22 2 p 1 m00 C m11 m22 q2 D sign .m02 m20 / 2 p 1 m00 m11 C m22 q3 D sign .m10 m01 / 2

q1 D sign .m21 m12 /

(5.50)

If a point (or a vector) P is first transformed by a quaternion Q1 and then by a quaternion Q2 , the resulting point (or vector) P0 is obtained by applying the transformation formula twice: P 0 D Q2 .Q1 P Q1 /Q2 D .Q2 Q1 /P .Q2 Q1 /

(5.51)

The above equation shows that the composite rotation is given by the quaternion product Q2 Q1 . Generalising this result, a series of rotational transformations performed using unit quaternions Q1 , Q2 , : : : Qk in that order, is equivalent to a single rotational transformation produced by the combined product quaternion (Qk : : : Q2 Q1 ).

5.5.2 Quaternions and Euler Angles In this section, we explore the relationship between unit quaternions and Euler angles. Using Eq. 5.44, we can represent elementary rotations about X, Y, and Z axes by angles , , respectively, as follows: QX D cos

2

; sin

2

; 0; 0

QY D cos ; 0; sin ; 0 2 2 QZ D cos ; 0; 0; sin 2 2

(5.52) (5.53) (5.54)

A sequence of Euler angle rotations f , , g is equivalent to the quaternion product QZ QY QX . We will denote this product by QE . Using the quaternion

92

5 Quaternions

multiplication rule in Eq. 5.10, we can easily express the components of QE in terms of the Euler angles. For convenience, the four quaternion components are arranged as a column vector in the equation below. cos cos cos C sin sin sin B 2 2 2 2 2 2 B B B B sin cos cos cos sin sin B 2 2 2 2 2 2 B QE D B B cos sin cos C sin cos sin B 2 2 2 2 2 2 B B @ cos cos sin sin sin cos 2 2 2 2 2 2 0

1 C C C C C C C (5.55) C C C C C A

Conversely, given a unit quaternion Q D (q0 , q1 , q2 , q3 ), we can compute the equivalent Euler angle representation by comparing the elements of the quaternion transformation matrix and the Euler angle transformation matrix. As an example, by equating the corresponding elements from only the first column and the third row of the matrices in Eqs. 5.47 and 5.30, we get the following expressions for the Euler angles , , in terms of quaternion components: 1

D tan

2 .q0 q1 C q2 q3 / 1 2q12 2q22

D sin1 .2q0 q2 2q1 q3 / 1 2 .q0 q3 C q1 q2 / D tan 1 2q22 2q32

(5.56)

There are many other ways in which the above parameters can be obtained by comparing the remaining elements of the two matrices. However, each derivation has its own set of singularities that need to be handled as special cases. For example, the unit quaternion 1 1 Q D p ; 0; p ; 0 2 2 presents a singularity for , with both the numerator and the denominator of the first equation in Eq. 5.56 becoming zero.

5.5.3 Negative Quaternion In this section, we consider another geometrical property of quaternions, taking QZ (Eq. 5.54) as an example. Figure 5.7 shows the plot of the first and the fourth nonzero components of QZ as the rotation angle is varied over two cycles from 0ı to 720ı .

5.6 Rotation Interpolation

93

Fig. 5.7 Plot showing the variation of quaternion components with rotation angle

Figure 5.7 shows that one cycle in quaternion space takes two revolutions in the Cartesian coordinate space. This means that two rotations by angles and 360 C that are geometrically equivalent, can have different quaternion representations. If a unit quaternion Q is given by Eq. 5.44, then replacing • with 360 C • we get, 360 C ı 360 C ı 360 C ı 360 C ı Q0 D cos ; l sin ; m sin ; n sin 2 2 2 2 ı ı ı ı D cos ; l sin ; m sin ; n sin 2 2 2 2 D Q

(5.57)

The above equation shows that both Q and Q represent the same rotational transformation. In the next section, we will consider the problem of interpolating between two orientations (which we had briefly touched on while introducing Euler angles), and then use some of the properties of quaternion rotations discussed above to define quaternion based interpolation methods.

5.6 Rotation Interpolation Animation sequences commonly use interpolated values between two poses. A pose defines the position and orientation an object. Position interpolation can be carried out either by interpolating between the corresponding coordinate values, or by fitting parametric curves (splines) through the points. However, interpolation between two orientations in three-dimensional space need not always produce a smooth transition from one orientation to another. Depending on the mechanism we use for representing rotations, we can get completely different interpolation sequences between the same initial and target orientations. Generally, one would prefer an interpolation that yields an optimal path that gives minimum rotation and uniform

94

5 Quaternions yB

Y

yB Y yB

O

Y

X

O O zB Z

X

X

zB

xB zB

Initial Configuration

Orientation-1

Orientation-2

Fig. 5.8 Initial configuration and two orientations of an object

angular velocity between two configurations. In this section, we will compare different interpolation methods using different representations of rotation we have considered so far, and establish that quaternions have a clear advantage over others. We define orientation as the result of a rotational transformation from the initial configuration of an object to its current configuration. A configuration is uniquely specified by an orthogonal system of axes fixed on the object. Some of these concepts are explained in a little more detail below with the help of an example. Figure 5.8 shows a simple model, “Hammer”, constructed using four primitives, a cylinder, a cone, a sphere and a cube. The figure also shows two orientations of this model. The initial configuration of the object defines its orientation when no rotational transformation is applied. In this configuration, an orthogonal right-handed system of body-fixed axes OxB yB zB coincides with the inertially fixed coordinate reference axes OXYZ. Without any loss of generality, we can assume that all rotations take place about the origin. The unit vectors along body fixed axes have components xB D (1, 0, 0), yB D (0, 1, 0), zB D (0, 0, 1) in the initial configuration. An orientation can be uniquely defined using the transformed components of these three vectors. For example, Orientation-1 in Fig. 5.8 is defined by the vectors xB D (0, 0, 1), yB D (1, 0, 0), zB D (0, 1, 0). During any rotational transformation the tips of these vectors move on a unit sphere centered at the origin (Fig. 5.9a). The rotational transformation of an object can thus be visualized using the trace of the unit vectors along the body-fixed axes on a unit sphere. Any unit vector has a spherical parameterization in terms of its azimuth (or longitude) ˛, and elevation (or latitude) ˇ (Fig. 5.9). The variation of the tip of a vector v D (xv , yv , zv ) on a unit sphere can be conveniently represented as a 2D-graph of the values (˛, ˇ) computed as follows: 1 xv ˛ D tan zv ! yv 1 ˇ D tan p (5.58) xv2 C z2v

5.6 Rotation Interpolation

95

a

b

Y

v

yB

zB

b X

O a Z

xB

Fig. 5.9 Spherical parameterization of rotations: (a) Movement of unit vectors attached to body axes during a rotation of the object. (b) Parametric representation of unit vectors on a sphere Table 5.1 Graph values (’,“) of the two orientations in Fig. 5.8

xB yB zB Transformation matrix

Orientation-1 (0, 0) (90, 0) (, 90) 2 3 0 1 0 4 0 0 1 5 1 0

0

Orientation-2 (˙180, 0) (, 90) (90, 0) 2 3 0 01 4 0 1 05 1 0 0

indicates an indeterminate value

We call the above method of representing the three-dimensional variations of a unit vector as the ˛ˇ-graph method. Note that when ˇ D ˙90ı , the value of ˛ is indeterminate. In the following sections, we will use ˛ˇ-graphs of the bodyfixed axes for a given interpolation sequence to compare the paths generated by different methods. For the example given in Fig. 5.8, the graph values (in degrees) of Orientation-1 and Orientation-2 are shown in Table 5.1. The variation of a graph between the two points will help us visualize how a sequence of rotational transformations operates on an object for transforming it from one orientation to the other. Another method for visualizing three-dimensional rotations is to show a small triangle (see Fig. 5.9a) at the position of one of the body axes (say, zB ) on the unit sphere, oriented towards another axis (say yB ). The triangle uniquely represents the three-dimensional orientation of the object. Triangles displayed at equal time intervals during a rotational transformation will clearly show the movement of an axis of interest, and also indicate the spin of the object about that axis (see Fig. 5.10).

5.6.1 Euler Angle Interpolation Let us first consider the interpolation between two orientations represented using Euler angles. For our example, we will use the Euler angle sequence f , , g

96

5 Quaternions

Fig. 5.10 Interpolation sequence generated using Euler angles (90, 90, 0) and (0, 90, 0)

introduced in Sect. 5.4.1. Given two sets of Euler angles f 1 , 1 , 1 g and f 2 , 2 , 2 g, all intermediate sets can be obtained using a linear interpolation between the corresponding Euler angles: D .1 t/

1

Ct

2

D .1 t/ 1 C t 2 D .1 t/1 C t2 ; 0 t 1:

(5.59)

The transformation matrix in Eq. 5.30 then defines the rotation from the initial configuration to the intermediate orientation. Earlier in Fig. 5.5, we saw examples of interpolation sequences generated in this manner. For the example given in Fig. 5.8, Orientation-1 is defined by Euler angles f 1 D 90, 1 D 90, 1 D 0g, and Orientation-2 by f 2 D 0, 2 D 90, 2 D 0g. The ˛ˇ-graph for the interpolation sequence is given in Fig. 5.10. For this specific example, linear interpolation in the domain of Euler angles also generates a perfect linear interpolation in ˛ˇspace, consisting of equidistant points. However, when we look at the trace of the hammer’s axis from Y direction to CX direction on the surface of the unit sphere, we observe that the rotational motion from the source to the destination in threedimensional space is not uniform. The “Hammer” example in Fig. 5.8 also presents an interesting aspect of Euler angles. Orientation-2 can have an infinite number of Euler angle representations given by f 2 D œ, 2 D 90, 2 D œg where œ is any value. Thus between the same two orientations, we can have several interpolation paths using Euler angles. As an example, the interpolated values obtained using œ D 170ı give a distinctly different and curvilinear path between Orientation-1 and Orientation-2, as shown in Fig. 5.11.

5.6.2 Axis-Angle Interpolation The equivalent angles and axes of rotation for both Orientation-1 and Orientation-2 can be computed from the corresponding transformation matrices using Eq. 5.41.

5.6 Rotation Interpolation

97

Fig. 5.11 Interpolation sequence generated using Euler angles (90, 90, 0) and (170, 90, 170)

Fig. 5.12 Interpolation sequence generated using the angle-axis transformation

The parameters for Orientation-1 are ı 1 D 120ı, l1 D 0.57735, m1 D 0.57735, n1 D 0.57735, and for Orientation-2 the values are ı 2 D 90ı , l2 D 0, m2 D 1, n2 D 0. A straightforward linear interpolation gives ı D .1 t/ı1 C tı2 l D .1 t/l1 C tl2 m D .1 t/m1 C tm2 n D .1 t/n1 C tn2 ; 0 t 1:

(5.60)

The interpolated vector will need to be normalized before constructing the transformation matrix in Eq. 5.38. The intermediate orientations generated using the above equation are shown in Fig. 5.12. In the example shown above, the angle axis transformation generates a nonuniform motion with a large variation in the angular velocity. As can be seen from both the ˛ˇ-graph and the trace on the sphere, the density of points around the source and the destination points is very large compared to the middle. The parameters used in the interpolation belong to completely different domains, the angle being a scalar and the axis of rotation being a vector. Quaternions help us to

98

5 Quaternions

Fig. 5.13 Interpolated and unit quaternions on a unit sphere in quaternion space

Q2

Q1

Interpolated quaternion

Unit quaternion

eliminate this disparity in the type of the interpolants, and achieve a rotation where both the axis of rotation as well as the rate of change of angle remain constant. In the next section, we consider a linear interpolation using quaternions.

5.6.3 Quaternion Linear Interpolation (LERP) Given two unit quaternions Q1 D fq0 (1) , q1 (1) , q2 (1) , q3 (1) g and Q2 D fq0 (2) , q1 (2) , q2 (2) , q3 (2) g, a linear interpolation gives the quaternion Q D .1 t/Q1 C tQ2 ;

0 t 1:

(5.61)

The quaternion resulting from the above equation is converted to a unit quaternion before a transformation of the form QPQ* is applied to all points P of the object. Every unit quaternion lies on a unit sphere in the four-dimensional space spanned by the quaternion basis (1, i, j, k). The interpolated quaternions obtained from Eq. 5.61 lie on a straight line between the two points Q1 and Q2 . Converting them to unit quaternions moves each interpolated quaternion to the surface of the sphere along a radial (Fig. 5.13), resulting in an uneven distribution of points and a corresponding non-uniformity in the angular velocity of the object. The speed in the middle of the interpolation path is generally much higher than the speed at the end points. The interpolated quaternions after normalization lie on an arc of a great circle between Q1 and Q2 . Continuing with our “Hammer” example in Fig. 5.8, the source and the target orientations in Table 5.1 can be converted into quaternions using Eq. 5.49. For Orientation-1, the quaternion parameters are q0 (1) D 0.5, q1 (1) D 0.5, q2 (1) D 0.5, q3 (1) D 0.5, and for Orientation-2, the values are q0 (2) D 0.71, q1 (2) D 0, q2 (2) D 0.71, q3 (2) D 0. The ˛ˇ-graph and the trace of the hammer axis on the sphere are shown in Fig. 5.14.

5.6 Rotation Interpolation

99

Fig. 5.14 Interpolation sequence generated using quaternion linear interpolation Fig. 5.15 Subdivision of the angle between unit quaternions

b

a

Q2

P

R

cosW

Q2

sinW

Q2′

Q

Q1

tW

Q1

The interpolation path obtained using quaternions is along a circular arc between the end points, which is often the most desired path. However, the non-uniform spacing of points along the arc indicates that the angular velocity is initially smaller, then increases towards the middle and slows down again towards the target.

5.6.4 Quaternion Spherical Linear Interpolation (SLERP) In the previous section we saw that linear interpolation generates intermediate quaternions along a chord between Q1 and Q2 (Fig. 5.13) on the unit sphere in quaternion space. If we subdivide the angle between Q1 and Q2 uniformly, then we will get an even distribution of points on the sphere. Such a distribution will also yield a smooth rotation of the object from one orientation to another with nearly constant angular velocity. The spherical linear interpolation (SLERP) technique uses this approach to compute intermediate quaternions. Figure 5.15 shows the geometrical constructions needed to derive the SLERP formula. In the figure, Q1 D fq0 (1) , q1 (1) , q2 (1) , q3 (1) g and Q2 D fq0 (2) , q1 (2) , q2 (2) , q3 (2) g are any two unit quaternions and P is another unit quaternion that is orthogonal to Q1 . Treating them as vectors in quaternion space, Q2 Q1 cos˝ is a vector

100

5 Quaternions

Fig. 5.16 Interpolation sequence generated using quaternion spherical linear interpolation

(denoted by R) from Q2 0 (the projection of Q2 on Q1 ) to Q2 , where ˝ is the angle between Q1 and Q2 . ˝ is computed from the following formula: cos ˝ D q0 .1/ q0 .2/ C q1 .1/ q1 .2/ C q2 .1/ q2 .2/ C q3 .1/ q3 .2/

(5.62)

Dividing R by its magnitude (sin˝), we get the unit quaternion in the direction of R. Thus P D

Q2 Q1 cos ˝ sin ˝

(5.63)

Figure 5.15 shows the angle between Q1 and Q2 subdivided using an interpolation parameter t (0 t 1), and the interpolated unit quaternion Q generated using this subdivision. Resolving Q along the orthogonal unit directions of Q1 and P we get Q D Q1 cos.t˝/ C P sin.t˝/

(5.64)

Substituting Eq. 5.63 and simplifying we get QD

Q1 sin ..1 t/˝/ C Q2 sin .t˝/ sin .˝/

(5.65)

The above equation has a singularity when ˝ D 0 or ˙180ı . When ˝ D 0, both the initial and final quaternions are the same, and therefore no interpolation is necessary. When ˝ D ˙180ı , Q2 D Q1 . From Eq. 5.57 we know that this condition also corresponds to the situation where both orientations are the same. The interpolated sequence generated by Eq. 5.65 for the “Hammer” example is shown in Fig. 5.16. Compared with the results obtained from previously discussed forms of interpolation, the smoothness of the interpolating curves as well as the uniformity in the distribution of points along them are noticeable. Spherical linear interpolation yields an optimal angle interpolation between two orientations with a constant axis of rotation. If the interpolation parameter is incremented in constant steps, spherical linear interpolation will generate a motion with constant angular velocity.

5.7 Quaternion Exponentiation Fig. 5.17 Two different interpolation paths on the quaternion sphere

101

a

b

Q1

Q2

-Q2

Q1

W W

Q2

When interpolating between two quaternions Q1 and Q2 , we have to make sure that the interpolation is performed along the shorter arc on the great circle through the two points on the quaternion sphere. If the angle ˝ between Q1 and Q2 is less than 90ı , we interpolate between the two quaternions (Fig. 5.17), otherwise we interpolate between Q1 and Q2 (Fig. 5.17). In other words, if Q1 • Q2 D cos˝ < 0, we negate the sign of Q2 . The value of cos˝ is computed using the formula in Eq. 5.62. The following sections discuss a few more applications of quaternions for representing transformations in a three-dimensional space.

5.7 Quaternion Exponentiation We will extend the notion of exponentiation from the field of complex numbers to the domain of quaternions and also define the associated logarithmic function that is consistent with exponentiation. However, there are some subtle differences between the way in which these operations are performed on real and complex numbers and the way they are applied to quaternions. From Eq. 5.6 we know that a unit complex number can be expressed as z D (cosı, sinı). The same complex number has an alternate representation in the form z D eiı . This is the well known Euler’s formula in complex numbers. We know that a unit quaternion can also be written as Q D (cosı, u sinı). Similar to complex numbers, an exponential notation for unit quaternions can be introduced as follows: Q D .cos ı; u sin ı/ D e uı

(5.66)

where u D (l, m, n) is a unit vector in three-dimensional space. For the time being, we will treat the above equation as only an alternate representation of unit quaternions. We will see the formal definition of the exponential function and how it is related to the above notation immediately after the next equation. The logarithm of the unit quaternion in Eq. 5.66 is defined as QL D log.Q/ D log.e uı / D .0; uı/ D .0; lı; mı; nı/

(5.67)

102

5 Quaternions

QL is a pure quaternion and its magnitude is ı. The definition of an exponential function for quaternions must be consistent with the above operation and the inverse of the log() function, such that exp(log(Q)) D Q. We thus have the following definition: exp.QL / D exp..0; uı// D .cos ı; u sin ı/ D e uı :

(5.68)

The above definition leads to the following important result for any unit quaternion Q D (cosı, u sinı), and any real value t: Qt D exp.tlog.Q// D exp..0; utı// D .cos.tı/; u sin.tı//

(5.69)

Note that the operations Qt and exp(QL ) both return unit quaternions. As a special case, when t D 0, we have Q0 D .1; 0; 0; 0/ for any unit quaternion Q:

(5.70)

Since quaternion multiplication is non-commutative, it immediately follows that Qa Qb ¤ Qb Qa and, log(PQ) ¤ log(P) C log(Q). However, the following equations are valid for all unit quaternions Q: Qa Qb D QaCb .Qa /b D Qab

(5.71)

We know that the unit quaternion Q given in Eq. 5.66 represents a rotation by an angle 2ı about the unit vector u passing through the origin. From Eq. 5.69, we see that raising Q to the power of t effectively changes the angle of rotation. Thus if 0 t 1, then Qt gives a unit quaternion that represents a partial rotation 2tı. This result is useful for interpolating between orientations. In the next section, we will define the relative quaternion between two orientations, and then apply Eq. 5.69 to perform incremental rotations along a path from the source orientation to the target orientation. As a result, we will get another equation for the quaternion spherical linear interpolation using the exponential notation.

5.8 Relative Quaternions In Sect. 5.6, we defined the three-dimensional orientation of an object using the parameters of rotation that transforms the object from its initial configuration to the current. This rotation can be represented by a unit quaternion. Thus two independent orientations of an object can be represented by two unit quaternions Q1 and Q2 (Fig. 5.18). In the following, we try to find the relative quaternion that performs

5.8 Relative Quaternions

103 Q2 Orientation-2

P2

Initial Configuration P Orientation-1

Relative quaternion

Q1

P1

Fig. 5.18 The relative quaternion transforms an object from one orientation to another

a rotation from the first orientation to the second. This relative quaternion can be easily obtained by noting how Q1 and Q2 transform points from one frame to another. In Fig. 5.18, the point P1 in Orientation-1 corresponds to the point P in the initial configuration. In other words, the quaternion Q1 transforms P into P1 . Similarly the quaternion Q2 transforms P into P2 in Orientation-2. Therefore, P1 D Q1 P Q1 P2 D Q2 P Q2

(5.72)

Now we seek a quaternion that transforms P1 into P2 . From the first equation above, we get the inverse transformation, P D Q1 P1 Q1

(5.73)

Substituting in the second equation, we have P2 D Q2 Q1 P1 Q1 Q2 D .Q2 Q1 /P1 .Q2 Q1 /

(5.74)

The above equation shows that the quaternion Q2 Q1 * transforms the point P1 into P2 , and therefore represents the transformation from Orientation-1 to Orientation-2. Note that Q2 Q1 * is a unit quaternion. Q2 Q1 * is called the relative quaternion between Q1 and Q2 . We now revisit the problem of interpolating between Orientation-1 and Orientation-2. Any intermediate orientation in the above example can be obtained by first applying the unit quaternion Q1 to get to Orientation-1 from the initial

104

5 Quaternions Listing 5.1 Pseudo-code for computing SLERP equation in exponent form

configuration, and then applying a partial rotation using the relative quaternion Q2 Q1 *. From Eq. 5.69, we know that this partial rotation can be effected by (Q2 Q1 *)t , where, 0 < t