Human pose estimation from depth image using visibility estimation and key points

Abstract

In this paper, we propose the upper body pose estimation algorithm using 3-dimensional model and depth image. The conventional ICP algorithm is modified by adding visibility estimation and key points - extreme points and elbow locations. The visibility estimation keeps occluded points from participating in pose estimation to alleviate the affection of self-occlusion problem. Introduction of extreme points and elbow locations, which are extracted using geodesic distance map and particle filter, improves the accuracy of pose estimation result. The optimal parameters of the model are obtained from nonlinear mathematical optimization solver. The experimental results show that the proposed method accurately estimates the various human poses with self-occlusion. © 2013 Springer-Verlag.