A NEW REPRESENTATION METHOD OF HEAD IMAGES FOR HEAD POSE ESTIMATION

In this paper, a discriminative representation method of head images is proposed, which is based on parts and poses for identity-independent head pose estimation. Head images are preprocessed to enhance the facial features and eliminate the identity information by skin color model and Laplacian of Gaussian transform. Then, the preprocessed images are used to construct a eigenpose subspace by a matrix factorization method. The testing head images are represented as the projections of the eigenpose subspace in which we can easily find the decision function for head pose estimation. The proposed representation method evaluated on the standard head pose database and real-time videos achieves higher pose estimation accuracy than other methods.