Abstract
In this thesis, we propose a 3D model-based method to track human body motion in monocular image sequences. The 3D human pose is obtained by fitting the projected contour of a 3D human model to the edge map in the video. The prior knowledge of human kinematics is integrated into the 3D human body model to reduce the pose search space and achieve reasonable pose estimation. The body segments of the 3D human model are modeled by generalized cylinders. We propose an efficient method to determine the projected contour of the 3D generalized cylinder model onto the image plane and handle the self-occlusion between body segments by identifying and eliminating the occluded contour. We apply one-way Hausdorff distance as the similarity measure for matching the projected contour with the edge map computed from video. With the 3D generalized cylinder human model, we can track and recover poses of human in a monocular image sequence through a hierarchical local pose search. Experimental results of applying the proposed 3D human body tracking algorithm to real video are given to show its performance.