Real-time head orientation estimation using neural networks

Abstract

Estimation of human head orientation is important for a number of applications such as human-computer interaction, teleconferencing, virtual reality, and 3D audio rendering. We present a system for estimating human head orientation based on visual information. Two neural networks are trained to approximate the functions that map an image of a head to the orientation of the head. We obtain ground-truth data for training and testing from an electromagnetic tracking device worn by subjects. Our experimental results demonstrate orientation accuracy within 10° with the subject free to move about at distances of three to ten feet from the camera. The system is designed to be robust to illumination changes and it runs in real time.