Near real-time three axis head pose estimation without training

Abstract

Head pose estimation methods evaluate the amount of head rotation according to two or three axes, aiming at optimizing the face acquisition process, or extracting neutral-pose frames from a video sequence. Most approaches to pose estimation exploits machine-learning techniques requiring a training phase on a large number of positive and negative examples. In this paper, a novel pose estimation method that exploits a quad-tree-based representation of facial features is described. The locations of a set of landmarks detected over the face image guide its subdivision into smaller and smaller quadrants based on the presence or lack of landmarks within each quadrant. The proposed pose descriptor is both effective and efficient, providing accurate yaw, pitch and roll axis estimates almost in real-time, without need for any training or previous knowledge about the subject. The experiments conducted on both the BIWI Kinect Head Pose Database and the challenging automated facial landmarks in the wild dataset, highlight a pose estimate precision exceeding the state-of-the-art with regard to methods not involving training and machine learning approaches.