Head Pose Recovery Using 3D Cross Model

This paper presents a new method for 3D head pose recovery from video by using a novel 3D cross model to track the head motion. Given an initial frontal face of a head, a biologically-plausible 3D cross model is constructed to represent the head which is robust to lighting and facial expression changes. By updating the model templates dynamically, the effects of self-occlusion, head large motion and gradual illumination changes can be diminished. When the head is hidden or moves out of scenes, the model is reinitialized automatically. Experimental results show that the method can efficiently recover the head pose and eliminate problems of the cumulative error and illumination changes.