Learning Non-linear Models of Shape and Motion

Abstract

Deformable models have been an active area of research in computer vision for a number of years. Their ability to model non-ridgid objects through the combination of geometry and physics has proven a valuable tool in image processing. More recently a class of deformable objects known as Point Distribution Models or Eigen Models have been introduced. These statistical models of deformation overcome some of the shortfalls of earlier deformable models by learning what is 'allowable' deformation, for an object class, from a training set of examples. This semi-automated learning procedure provides a more generic approach to object recognition, tracking and classification. Their strength lies in their simplicity and speed of operation, allowing the robust ability to model complex deformations in cluttered environments. However, the automated construction of such models leads to a breakdown of the fundamental assumptions upon which they are based. Primarily, that the underlying mathematical model is linear in nature. Furthermore, as more complex objects are considered, these assumptions fail completely and what is produced is an unreliable model.

This work addresses these problems and presents novel techniques for the automated construction and application of non-linear deformable models, which retain the speed, and simplicity of the linear Point Distribution Model. It is further shown how these non-linear models can be augmented with probabilistic temporal constraints, which are essential in object tracking and classification.

This work presents, in essence, three developments to the field. Firstly, a piecewise linear approach to modelling non-linearity is proposed and results demonstrated that show its accuracy in modelling both low and high dimensional datasets with heavy non-linearity. The technique is then extended to the automated construction of models. Secondly, it is shown how the piecewise approach can be augmented with temporal constraints and used in both model prediction, animation and for the support of multiple hypotheses during tracking. It is further shown how these temporal models can be extended to incorporate information from other sources, providing more reliable tracking in the absence of complete training data. Thirdly, it is shown how elements can be combined statistically and used to infer information about an object from its shape alone. Using human motion capture as an example, it is demonstrated that models can be assembled which allow 3D structural information about body pose and motion to be inferred from a monoscopic image sequence using only natural features of the body as markers.


Welcome to my thesis download page.

The thesis is available in pdf as a complete document or as individual chapters.

It uses colour extensively throughout so if you are planning on printing in black and white some of the figures may not be too clear.

Page Title Download
1-249 Full Thesis Learning non-linear Models of Shape and Motion 9.232KB PDF

 

Page Chapter Title Download
i-xi

Intro Pages

Title Page
Abstract
Acknowledgements
Declaration
Table of Contents
Table of Figures
Abreviations
82KB PDF
1-4 Chapter 1 Introduction 55KB  PDF
5-14 Chapter 2 Literature Review 75KB  PDF
15-36 Chapter 3 Linear Point Distribution Models 638KB  PDF
37-58 Chapter 4 Enhancing Tracking Using Colour 873KB  PDF
59-89 Chapter 5 Cluster Based non-linear Point Distribution Models 594KB  PDF
90-113 Chapter 6 Cluster Constraints on Shape Space 304KB  PDF
114-153 Chapter 7 Adding Temoral Constraints 1,119KB  PDF
154-171 Chapter 8 3D Point Distribution Models 2,608KB  PDF
172-193 Chapter 9 Extending the Point Distribution Model 595KB  PDF
194-198 Chapter 10 Closing Discussion 55KB  PDF
199-203 Appendix A k-means and Fuzzy k-means Clustering 42KB  PDF
204-225 Appendix B Volumetric Segmentation 423KB  PDF
226-237 References   31KB  PDF

This document and its parts are covered by copyright and no repoduction of the document or content may be made without the permission of the author.
Any reference to this work must contain an appropriate reference to the author.

Richard Bowden 2000