Motion tracking of 3D articulated figure from 2D information of a single video stream

Abstract

A Huge amount of human activities and motions are recorded in a single sequence of uncalibrated video images. In this thesis, we present a method for reconstructing the 3D human motions with respect to 2D motion information contained in a single video stream. 2D information is formed as a sequence of 2D position set and acquired by interactive marking. The reconstruction process can be formulated as an inverse kinematics problem. We model a human body as a non-linear dynamic system and adopt a extended Kalman filter as a inverse kinematics solver so as to resolve depth ambiguity. Our Kalman filter tracks the joint motion spanning entire orientation space by parameterizing the orientations incrementally with rotation vector. We augment the measurement model with the angle-bound constraints prevent infeasible motions. To avoid singularities which arise in inverse kinematics problem, the analyses on the Singularity-Robust inverse are adapted to make the Kalman filter avoid singular states.