Saliency model based head pose estimation by sparse optical flow

Head pose plays an important role in Human-Computer interaction, and its estimation is a challenge problem compared to face detection and recognition in computer vision. In this paper, a novel and efficient method is proposed to estimate head pose in real-time video sequences. A saliency model based segmentation method is used not only to extract feature points of face, but also to update and rectify the location of feature points when missing happened. This step also gives a benchmark for vector generation in pose estimation. In subsequent frames feature points will be tracked by sparse optical flow method and head pose can be determined from vectors generated by feature points between successive frames. Via a voting scheme, these vectors with angle and length can give a robust estimation of the head pose. Compared with other methods, annotated training data set and training procedure is not essential in our method. Initialization and re-initialization can be done automatically and are robust for profile head pose. Experimental results show an efficient and robust estimation of the head pose.