A graphics processing method for occlusion reasoning in monocular tracking using nonparametric belief propagation

Tracking of articulated subjects like humans is plagued by complex occlusions among the interacting parts of the subject. The use of monocular views only accentuates this problem. We propose a method to jointly estimate a field of pixel labels along with an articulated pose within the framework of nonparametric belief propagation. We show how this algorithm may be efficiently implemented using a graphics processing unit to streamline complex occlusion reasoning computations via graphics rendering.