Head Pose Estimation Via Sparse Manifold

Statistics indicate that approximately75%of the faces in photographs are non-frontal. For face recognition and related issues, pose estimation is an importantpretreatment process. In the past decade, the research on face recognition has made greatprogress, but robust face recognition is difficult when the pose changes. Since the accurateestimation of pose is a prerequisite to solve these problems, the pose estimation hasattracted more and more attention. And the key problem of accurate head pose estimationis that complicated3D structure of face and multi-variated situation of capture.As the input samples are usually wrong registration level. We propose one methodbased on sparse representation and another method based on template matching. The twomethods do fast detection on detected face area, cut and fill the extra background area. Soas to enlarge the relative face area, improve the registration level between the two datasets,which provides better input data for the following feature extraction and classification.Additional, how to design an efficient classifier is a key problem in the poseestimation methods based on appearance features. As pose estimation is also regard as aproblem of face vision perception. In this paper, we try to extend the sparse representationtheory to the pose estimation problem, and study the topic that how to dynamicallygenerate better overcomplete dictionary dataset. Also we want to improve thediscriminant ability and regression ability of sparse representation by use of thecontinuous information of head poses.