Human Pose Estimation in Stereo Images

In this paper, we address the problem of 3D human body pose estimation from depth images acquired by a stereo camera. Compared to the Kinect sensor, stereo cameras work outdoors having a much higher operational range, but produce noisier data. In order to deal with such data, we propose a framework for 3D human pose estimation that relies on random forests. The first contribution is a novel grid-based shape descriptor robust to noisy stereo data that can be used by any classifier. The second contribution is a two step classification procedure, first classifying the body orientation, then proceeding with determining the full 3D pose within this orientation cluster. To validate our method, we introduce a dataset recorded with a stereo camera synchronized with an optical motion capture system that provides ground truth human body poses.