Efficient Estimation of Human Upper Body Pose in Static Depth Images

Automatic estimation of human pose has long been a goal of computer vision, to which a solution would have a wide range of applications. In this paper we formulate the pose estimation task within a regression and Hough voting framework to predict 2D joint locations from depth data captured by a consumer depth camera. In our approach the offset from each pixel to the location of each joint is predicted directly using random regression forests. The predictions are accumulated in Hough images which are treated as likelihood distributions where maxima correspond to joint location hypotheses. Our approach is evaluated on a publicly available dataset with good results.