Abstract: Abstract: Real-time head pose estimation plays a crucial role in the application of human-computer interaction and face analysis, but accurate head pose estimation methods still face certain challenges. In order to improve the accuracy and robustness of the head pose estimation, this paper combines the geometry-based method and the learning-based method for head pose estimate. On the basis of face detection and face alignment, the geometric feature of the color image and the local area depth feature of the depth image are extracted, combining with the normal and curvature feature of the depth block to form the feature vector group, and then the random forest method is used to do the training. Finally, all decision trees are involved in the vote, and the resulting Gaussian distribution of the head pose is filtered by thresholds to further improve the model’s accuracy. Experimental results show that the proposed method has higher accuracy and robustness than the existing head pose estimation methods.

Key words: Keywords: head pose estimation, random forest, RGBD data, geometric feature, depth feature