Model-Based High-Dimensional Pose Estimation with Application to Hand Tracking

Abstract

This thesis presents novel techniques for computer vision based full-DOF human hand motion estimation. Our main contributions are: A robust skin color estimation approach; A novel resolution-independent and memory efficient representation of hand pose silhouettes, which allows us to compute area-based similarity measures in near-constant time; A set of new segmentation-based similarity measures; A new class of similarity measures that work for nearly arbitrary input modalities; A novel edge-based similarity measure that avoids any problematic thresholding or discretizations and can be computed very efficiently in Fourier space; A template hierarchy to minimize the number of similarity computations needed for finding the most likely hand pose observed; And finally, a novel image space search method, which we naturally combine with our hierarchy. Consequently, matching can efficiently be formulated as a simultaneous template tree traversal and function maximization.