Supervised Local Subspace Learning For Head Pose Estimation

Head pose estimation is one of the usual means for nonverbal communication,and plays an important role in human daily life. In the field of computer vision, facepose estimation can be regarded as a series of image processing operations whichaim to extract effective features and to infer the direction of the face. Nowadays,face pose estimation had been widely applied to many fields, e.g., human-computerinteraction, face recognition, and driver assistance, etc.In this thesis, we develop a face pose estimation system, which includes threemajor modules: face detection, face tracking, and pose estimation. First, we providea brief survey on the face detection and face tracking methods, and descrive the facedetection and tracking algorithms used in the proposed system.Second, we describe the supervised local subspace learning (SL~2) method forface pose estimation. Specifically, we introduce the basic idea and the loss funtionsof the SL~2algorithm. From the aspect of manifold learning, we model the training ofthe face pose estimation algorithm as a problem of estimating the mean and basis ofthe local subspace. Compared with other methods,SL~2can achieve higheraccuracy and is more robust against the imbalance in training set. Besides, wefurther proposed a closed form solution forSL~2, which can significantly improvethe computational efficiency during the training stage.Finally, we present our online face pose estimation system by introducing thebuilding process of the system, the display of the interface, description of theseveral major modules,, and some discussions and analysis on the performance ofthe system.