Identity- and illumination-robust head pose estimation using manifold learning

Head pose estimation using manifold learning is challenging due to other appearance variations such as identity and illumination changes. To address the problem, we propose to incorporate supervised information (pose angles of training samples) into the process of manifold learning. Most manifold learning algorithms have two common variables: inter-point distances and graph weights, which can greatly affect the property of the constructed manifold. We propose to redefine these variables by constraining them with the pose angle information. In addition, since the environmental illuminations are distributed in the low-frequency component of the image and the texture-based feature is irrelevant to the pose variation, we use the proposed Localized Edge Orientation Histogram (LEOH) rather than the pixel intensity feature for manifold learning. The experimental results show that our method has the highest estimating accuracy and is robust to identity and illumination.