Recovery of 3D articulated motion from 2D correspondences

Abstract

Recovering the 3D motion of the human body is an important problem in computer vision. Applications that would benefit from 3D motion include physical therapy, computer user interfaces, and 3D animation. Unfortunately, recovering 3D position from one 2D camera is an inherently ill-posed problem. This thesis focuses on recovery of 3D motion of an articulated model using 2D correspondences from an existing 2D tracker. A number of constraints are used to aid in reconstruction: (i) kinematic constraints from a 3D kinematic model, (ii) joint angle limits, (iii) dynamic smoothing, and (iv) key frames. These methods are used successfully to recover 3D motion from video sequences. Also presented is a method for recovering 3D motion from motion capture data, as well as a method for recovering kinematic model connectivity from 2D tracks.