Model based tracking of articulated objects
Date of Issue2014
School of Computer Engineering
Human motion capture (HMC) is an active area of research in the computer vision community. The challenging nature of the problem and the diverse application potential has made human pose tracking, which is a fundamental problem for HMC systems, an interesting research area. The approaches to human pose tracking are divided as model free , model based and bottom-up approaches in the current literature. This thesis contributes to model based approach to human pose tracking. Though the state-of-the-art systems in this approach have shown high quality results for human motion reconstruction, they incur a high overhead to achieve this. Hence designing efficient techniques that achieve the same results at an acceptable computational cost has become a crucial research problem, which is addressed in this thesis. In order to achieve this, three different approaches are pursued in this thesis. We propose a stochastic search method referred to as parametric annealing (PA) for human pose inference. PA improves on the current approaches by reusing function evaluations across annealing layers combined with an appropriate annealing schedule. This results in an efficient inference technique. Furthermore, it relies on a parametric form that enables a more accurate estimation of the latent state. We also present an improved likelihood formulation for HMC in this thesis. The current methods to HMC use silhouettes and appearance to infer the human pose, which incur a significant overhead due to rendering and matching. We show that by using directional chamfer matching, better tracking results are obtained using significantly lower overhead. Furthermore, we show that the multi-view scenario in which HMC systems operate provide unique opportunities that can be exploited to estimate key parameters of the system. We further present a technique that enables hierarchical decomposition of the articulated human model. The current systems are not efficient due to the high dimensionality of the state space required to express a human pose. We show that the high dimensional state space can be decomposed into many smaller dimensional subspaces. Furthermore, we describe an efficient inference scheme that makes use of the decomposed subspaces and achieve state-of-the-art results using a small fraction of the computational resources. The inference scheme we describe is a general optimization procedure that can exploit the structure in the problem and result in efficient optimization.