Monocular Human Motion Tracking with Non-Connected Body Part Dependency

Abstract

2D articulated human pose tracking in monocular image sequences remains an extremely challenging task due to background cluttering, variation in body appearance, occlusion and imaging conditions. Most of the current approaches only deal with simple appearance and . Tadjacent. T or connected body part dependencies, especially the tree-structured priors assumed over body part connections. Such prior excludes the dependencies between . Tnon\small{-} connected. T body parts which could actually contribute to tracking accuracies. Building on the successful pictorial structures model, we propose a novel framework for human pose tracking including more dependencies of . Tnon\small{-}connected. T body parts. In order to implement inference efficiently for the proposed model, we introduce a factor graph to factorize all the unary term and all dependencies that are modelled in the pairwise term of the proposed model. In this paper, we also observe that the posterior marginals of each part from the tree structure model satisfy a Gaussian distribution. Based on this property, the sampling procedure becomes straight-forward and the search space can be shrunk effectively. We incorporate a simple motion constraint to capture the temporal continuity of body parts between frames, since the positions/orientations of body parts usually change smoothly between consecutive frames. In addition, we introduce a full body detector as the first step of our framework to reduce the search space for pose tracking. We also exploit the temporal continuity of body parts between frames by incorporating constraints on the location distance and the orientation difference for each body part between two successive frames. We evaluate our framework on two challenging image sequences and conduct a series of experiments to compare the performance with the approaches based on the tree-based model. The results illustrate that the proposed framework improves the performance significantly.