Real time head tracking of depth sequence with error recovery

Generally head tracking is based on the texture of face. The appearance template methods (skin-color-depended or texture) tracking results largely depend on a statistical color models obtained by observed image set for initialization. However, it remains an open problem on how to accurately initialize of models according to large variability in appearance of human with different clothing altering in appearance due to movement over time. It challenged tracking process with texture and illumination problem. This paper presents a real-time solution for head tracking without sensitive to texture and illumination, which based on sequence captured by depth sensor can handle the extreme rotation e.g. face totally turning backward. And the solution has the advantage of preliminary tracking error detecting and self error recovery.