Human Pose Tracking from Monocular Image Sequences

Abstract

This thesis proposes various novel approaches for improving the performance of automatic 2D human pose tracking system including multi-scale strategy, mid-level spatial dependencies to constrain more relations of multiple body parts, additional constraints between symmetric body parts and the left/right confusion correction by a head orientation estimator. These proposed approaches are employed to develop a complete human pose tracking system. The experimental results demonstrate significant improvements of all the proposed approaches towards accuracy and efficiency.