Head pose estimation and its application in TV viewers' behavior analysis

As a reliable indicator of visual gaze direction, head pose implies a person's visual attention and interest. Therefore, head pose information extracted from face images serves as important input in many applications. In this thesis, a coarse-to-fine head pose estimation method is proposed, by decomposing the original pose space in a hierarchical structure. The estimation begins with a coarse step to identify a subspace that encompasses a set of head pose candidates. Then a subsequent fine estimation is conducted within the subspace, generating a refined result. Besides, to eliminate irrelevant information within a face image, we propose to detect Region of Interest (ROI) by exploring importance degree of image points. Furthermore, we build an application of analyzing TV viewers' behaviors from video recordings, by integrating face detection, face tracking and head pose estimation. Based on head pose and face motion, a viewer's behavior is identified to be focused or unfocused.