Head pose Estimation via Direction-sensitive Feature and Random Regression Forests

Estimating the head pose is still a unique challenge for computer vision system. Previous methods at solving this problem have often proposed solutions formulated in a classification setting. In this paper, we formulate pose estimation as a regression problem to achieve robustness. We propose to use gradient orientation histograms based random regression forests for the task. Firstly, each sample image is divided into overlapped patches, and direction-sensitive features of patches are extracted. Then we train a random regression forest on these patches. Experiments are carried out on public available database, and the result shows that the proposed algorithm outperforms some other approaches in both accuracy and computational efficiency.