Markerless Human Motion Tracking from Monocular Videos

In this paper, a novel approach is proposed for tracking markerless human motion in monocular videos to capture the articulate motion data. With an articulated human model constructed, the new approach uses the probability density propagation of the particle filters through the learnt motion model and likelihood computing with the appearance models to track the human motion. The method is capable of automatically recovering from tracking failures. It can also process the occlusion and auto-occlusion problem correctly. Experimental results from real monocular videos show that the new approach is robust and the tracking results are satisfactory.