Head-Pose Estimation Based on Lateral Canthus Localizations in 2-D Images

Abstract

Head-pose estimation plays an important role in computer vision. The head-pose estimation aims to determine the orientation of a human head by representing the yaw, pitch, and roll angles. Implementations can be achieved by different techniques depending on the type of input and training data. This article presents a simple three-dimensional (3-D) face model for estimating head poses. The personalized 3-D face model is constructed by 2-D face photographs. A frontal face photograph determines the plane coordinates of facial features. By knowing the yaw angles in the other averted face photograph, the depth coordinates can be determined. The yaw angle of the averted face is evaluated by the canthus positions. Once the 3-D face model is constructed, we can find the matching angles for a target head pose in a query 2-D photograph. The personalized 3-D face model rotates itself about the x -, y -, and z -axes and then projects its facial features onto plane coordinates. If the rotation angles are correct, the disparities between the 2-D facial features and those in the query face photograph are supposed to be at their minimum. The personalized 3-D face model is validated with the University of South Florida human-identification database. The performance of the proposed head-pose estimation is evaluated on the Biwi Kinect head-pose database and Pointing’04 head-pose image database. The results show that the proposed method outperforms state-of-the-art technologies on both benchmark databases.