Independent Head Pose Estimation Method Based On Manifold Learning Identity Research

With the development of information society, the artificial intelligence related application demand is increasing, more and more information technology involves the application of face relevant, such as face recognition, 3D face model initialization, face tracking, facial image quality assessment, virtual reality and intelligent interaction, and head pose estimation is a key technology and one of the necessary steps to solve these problems. It has important research significance and practical value. In recent years, it becomes a very important research subject to pattern recognition and computer vision, and has attracted more and more attention.This paper focus on head pose estimation problem. By analyzing and comparing the advantages and disadvantages of the existing head pose estimation approaches. We study manifold learning method which is more suitable for solving the person-independent head pose estimation.First, we analyze the existing manifold learning algorithms by the corresponding experiment results for head pose estimation, and found that there are two disadvantages for traditional manifold learning algorithms to estimate head pose: neither using class labels of head image samples nor making use of the spatial information of image pixels.Second, local linear embedding (LLE) is a classical manifold learning algorithm, it can compute fast and need few parameters, and it is suitable for large scale high dimensional data processing. But there are two disadvantages mentioned above by directly using traditional unsupervised LLE to solve head pose estimation problem. Therefore, we combined biased manifold learning approach with image euclidean distance to improve LLE. When embedded with general regression neural network (GRNN), the poses of test samples are estimated with multivariate linear regression. The comparative experiments on FacePix database shows that the proposed method gets better head pose estimation.