Head pose estimation has been widely studied in recent decades due to many significant applications. Different from most of the current methods which utilize face models to estimate head position, we develop a relative homography transformation based algorithm which is robust to the large scale change of the head. In the proposed method, salient Harris corners are detected on a face, and local binary pattern features are extracted around each of the corners. And then, relative homography transformation is calculated by using RANSAC optimization algorithm, which applies homography to a region of interest (ROI) on an image and calculates the transformation of a planar object moving in the scene relative to a virtual camera. By doing so, the face center initialized in the first frame will be tracked frame by frame. Meanwhile, a head shoulder model based Chamfer matching method is proposed to estimate the head centroid. With the face center and the detected head centroid, the head pose is estimated. The experiments show the effectiveness and robustness of the proposed algorithm.
Deep convolutional neural network-based Bernoulli heatmap for head pose estimation