Head Pose Estimation Based on Tree-structure Cascaded Random Forests in Unconstrained Environment

Head pose estimation is an important evaluating indicator of human attention, which depends on many factors, such as illumination, noise, identification, occlusion and so on. In order to enhance estimation efficiency and accuracy, this paper presents tree-structure cascaded random forests to estimate head pose in different quality images. First, in order to eliminate the influence of different environment noise, combined texture features in random forests for positive facial patch classification are extracted, which will be the privileged inputs to estimate head pose. Second, a coarse-to-fine approach is proposed to estimate head pose both in the yaw and pitch, which is called tree-structure cascaded random forests. Third, an adaptive Gaussian mixture model is used to enhance discriminate vote energy in the tree distribution. This framework is evaluated in unconstrained environmental datasets. The experiments show that the proposed approach has a remarkable and robust performance in different quality images.