A novel hybrid approach to upper-body human motion capture

This paper describes an upper-body human motion capture system which combines a learning-based algorithm for human torso posture recognition with a model-based approach which estimates the human arms pose. The system uses a skin colour tracker to capture the movement of hands and face, and stereo data to obtain the silhouette of the human. The disparity map generated by the stereo vision system is also employed to perform a learning-based stage, providing an estimation of the torso posture. Experimental results show that the proposed approach runs at 20 frames per second in natural scenarios where the human is not required to wear special markers.