Pose Estimation Based On Pose Cluster And Candidates Recombination

Pose estimation in static images is a middle level vision problem whose purpose is to detect the body parts such as head, torso, limbs, etc. and output the position and angle parameters of the body parts. In the estimation process, we need to utilize both the low level color, boundary, gradient image characteristics and the relations of human body parts distribution such as geometric visual and kinetics constraints. The task is still challenging because of the wide changes of human poses, self-occlusion, various body color and size and background interference. As the result of pose estimation can be the basic information for computer to understanding and analysis human behavior, pose estimation has wide application prospects in intelligent monitoring, man-machine interaction and human motion capture.In this paper, we propose a pose estimation algorithm based on pose cluster and body part candidates recombination. Different from the previous single global PS model method, in order to increase the distinctiveness of body part detector meanwhile make the model be suitable for various poses, our approach clusters the poses before modeling, which makes the poses within each cluster have similar part appearances and spatial location relation. And the clustered mixture PS model can handle various poses based on the clusters. Multiple linear regression algorithm is employed to weighted average the best detection probability of each pose cluster to determine which cluster the test pose belongs to. In addition, we extract some of the best estimation results in the optimal clustered model as the candidates of body parts. And then we recombine them by constructing cost function and constraints on the candidates to overcome the individual body part false detection problem and double-counting problem. Finally, we obtain the optimal pose estimation by integer linear programming.Experiments on the publicly challenging LSP dataset show that the detection accuracies of each part and overall pose estimation are improved significantly in our approach comparing with the existing methods. The experiments also proved the effectiveness of pose cluster and that the double-counting problem can be overcome to some extent with the constraints we proposed.