Articulated body tracking using dynamic belief propagation

An efficient articulated body tracking algorithm is proposed in this paper. Due to the high dimensionality of human-body motion, current articulated tracking algorithms based on sampling [1], belief propagation (BP) [2], or non-parametric belief propagation (NBP) [3], are very slow. To accelerate the articulated tracking algorithm, we adapted belief propagation according to the dynamics of articulated human motion. The searching space is selected according to the prediction based on human motion dynamics and current body-configuration estimation. The searching space of the dynamic BP tracker is much smaller than the one of traditional BP tracker [2] and the dynamic BP need not the slow Gibbs sampler used in NBP [3-5]. Based on a graphical model similar to the pictorial structure [6] or loose-limbed model [3], the proposed efficient, dynamic BP is carried out to find the MAP of the body configuration. The experiments on tracking the body movement in meeting scenario show robustness and efficiency of the proposed algorithm.