Regression Forests for Head Pose Estimation Analysis

In order to tracking head pose motion of pilots in flight simulator,new and stable algorithms of head pose estimation are imperative and depth data are considered to be used to estimate head pose.We implement a random-regression-forest-based framework for head pose estimation.This framework uses massive human face scan model database annotated with head position and orientation,and patches are sampled randomly and sent to random forest for training.After training parts of leaf nodes save the Gaussian distributions for head position and orientation.Consequently head pose estimation is converted to searching for votable patches from test data,and votes of these patches are used to estimate head pose parameters.We analyze and demonstrate the effect of random forests' parameters and image features.According to experiments the approach can handle real data when large and rapid head rotations,partial occlusions,and facial expressions exist.