Effective Pose Estimation Using Pictorial Structure Model with Constraints of Appearance and Relative Position Based on Symmetry Relations

Abstract

In this paper, we propose an efficient method for re-estimating a human pose by applying constraints to a tree-structured Pictorial Structure (PS) model in pose estimation. The PS model is not robust to double-counting and disturbing textures such as background, hence our proposed method consists of the following two key ideas. (1) Applying an appearance constraint using the similarity of color distribution for a pair of symmetric parts to be robust against background. Then, if the pair of symmetric parts is estimated as a same position, the likelihood of the constraint is higher, and the constraint leads to the double-counting problem, and so an auxiliary constraint based on distance of the symmetric parts is applied. (2) To suppress unnatural incorrect pose estimation by using constraint based on human natural poses as prior knowledge. Here, clusters that represent various kinds of poses are used; however, the attribution cluster of a test image is obtained from a single PS model to reduce the computational cost. In the experiment, the proposed method reaches up to 80% in Percentage of Correct Parts (PCP), which outperforms the conventional PS models. In addition, the computation time of the proposed method is 1.25 times than the original PS model. Besides, the proposed method is a high speed and performance compared with the other improved method.