Dirichlet-tree cascaded Hough forests for continuous head pose estimation

We propose a hierarchical regression approach, Dirichlet-tree cascaded Hough forests (DCHF), which is based on deep learning for continuous head pose estimation in unconstrained environment, e.g., poses, illumination, occlusion, low image resolution, expressions and make-up. First, positive facial patches are learned and extracted from facial area to eliminate the influence of noise. Then, in order to estimate continuous head pose efficiently, multiple probability models are learned in four layers of the DCHF, i.e., the patch´s classification, the head pose angles, and offset probabilities mapping in the Hough space in a hierarchical way. Moreover, our algorithm takes a weighted and cascaded Hough voting method, where each positive patch extracted from the face can cast the efficient vote for head pose estimation. Experimental results on different public databases demonstrate the robustness and accuracy of the proposed approach to continuous head pose estimation.