Model-Based Head Pose Tracking With Stereovision

We present a robust model-based stereo head tracking algorithm that operates in real time on a commodity PC. The use of an individualized three-dimensional head model, coupled with the epipolar constraint from the stereo image pair, greatly improves the robustness of the tracking. Experimental results have shown that our method is able to track all the six degrees of freedom of the rigid part of head motions, over extended period of time, in the presence of large angular and translational head motions, partial occlusions, and/or dramatic facial expression changes. Applications include human-computer interaction and eye-gaze correction for video conferencing.