3D human body tracking in unconstrained scenes

Abstract

The 3D human body tracking from videos in unconstrained scenes is a challenging problem and has widespread applications. In this paper, we introduce a novel framework that incorporates the graph-based human limbs detection into the articulated Bayesian tracking. The 3D human body model with a hierarchical tree structure can describe human's movement by setting relevant parameters. Particle filter, which is the optimal Bayesian estimation, is used to predict the state of the 3D human pose. In order to compute the likelihood of particles, the pictorial structure model is introduced to detect the human body limbs from monocular uncalibrated images. Then the detected articulated body limbs are matched with each particle using shape contexts. Thus the 3D pose is recovered using a weighted sum of matching costs of all particles. Experimental results show our algorithm can accurately track the walking poses on very long video sequences. © 2009 Springer-Verlag Berlin Heidelberg.