Robust 3D head tracking by online feature registration

This paper presents a robust method for tracking the po- sition and orientation of a head in videos. The proposed method can overcome occlusions and divergence problems. We introduce an online registration technique to detect and register feature point of the head while tracking. A set of point features is registered and updated for each reference pose serving a multi-view head detector. The online fea- ture registration rectifies error accumulation and provides fast recovery after occlusion has ended, while preventing di- vergence problem which frequently occurs in conventional frame-to-frame tracking methods. The robustness of the proposed tracker is experimentally shown with video se- quences that include occlusions and large pose variations.