Human pose tracking based on partitioned sampling particle filter and multiple cues fusion

In this paper, we develop a method for tracking markless human pose in monocular video sequences. The number of required particles will grow exponentially when particle filter is applied to high dimensional tracking problems such as tracking human body poses, and particle filter with partitioned sampling is adopted to deal with this problem. We design a 2D human body model with constraints, and put forward a new adaptive way for fusing color, edge and motion cues together to construct the weighing function of particles. When calculating the likelihood function for each particle, we adaptively choose different templates and features according to the occlusion relationship among correlated body limbs. Thus, the proposed algorithm is capable of dealing with complicated occlusions among body limbs. In addition, we introduce a simplified belief propagation (BP) method to propagate the weights of limb observations to the corresponding particles along the edges of the body model, which can make a set of particles carry multiple constraints. Then we test the method on three video sequences which contain heavy background occlusion, complex human motion and self-occlusion, and the experimental results show that our method is effective and robust.