Head Pose Estimation Using Covariance of Oriented Gradients

Traditional appearance-based head pose estimation methods use the holistic face appearance as input and then employ statistic learning methods to extract low dimension features for classification. However, the face appearance may be more related to the unique identity of an individual rather than head poses. In this paper, we propose an image descriptor, covariance of oriented gradients (COG), for head pose estimation. This descriptor computes the covariance matrix of gradient based image features which characterizes the geometry structure of head pose images. To incorporate the spatial information, a head image is divided to several cells and the covariance matrices of all the cells are combined as the original image descriptor. Under the LogEuclidean metric, the original image descriptor is mapped into vector space, and then linear discriminant analysis is employed to find the discriminative low dimensional features. Experiments show that the proposed method outperforms two other state-of-the-art methods in terms of estimation accuracy and robustness on image resolutions.