Head Pose Estimation Based on Manifold Embedding and Distance Metric Learning

In this paper, we propose an embedding method to seek an optimal low-dimensional manifold describing the intrinsical pose variations and to provide an identity-independent head pose estimator. In order to handle the appearance variations caused by identity, we use a learned Mahalanobis distance to seek optimal subjects with similar manifold to construct the embedding. Then, we propose a new smooth and discriminative embedding method supervised by both pose and identity information. To estimate pose of a head new image, we first find its k-nearest neighbors of different subjects, and then embed it into the manifold of the subjects to estimate the pose angle. The empirical study on the standard databases demonstrates that the proposed method achieves high pose estimation accuracy.