Probabilistic Labeling of Human Body Parts

J. Ben-Arie, D. Sivalingam, and S. Rajaram (USA)


Human Body Part Labeling, Optimal Hypothesis Estima tion, Spatial Distributions of Torso Center


We develop a novel method for human body parts labeling, given a set of body parts that could be represented by bars in the image. First, we assume that in addition to the head and torso, the human body has 8 major body parts namely, the upper and lower arms (left and right) and the upper and lower legs (left and right) and all these parts can be ap proximated by cylinders that appear as rectangular bars in the image. One of the orginal contributions of this paper is the novel use of the spatial distributions of the location and orientation of the torso employed to find a correct role assignment for each bar by maximizing a joint probability function. Thus, we formulate the problem of labeling hu man body parts as an optimal hypothesis selection problem where each hypothesis corresponds to a particular role as signment. Next, we develop an algorithm which identifies the optimal hypothesis. The performance of our proposed approach is evaluated by testing it with 90 images depicting complex human poses and we obtain robust results (correct labeling in 98.7 % of the tested cases). We also test our method in cases of bar detection errors, occlusion and noisy images and obtain quite good results.

Important Links:

Go Back