3D head pose estimation using adaptive face template

Abstract

The estimation of human head pose has become an important issue in human-computer interaction field. Several real-world applications of head pose estimation have been employed, such as a virtual avatar control and car driver assistance system, and the usage and demand keeps growing in the field. In the last decade, various approaches have been attempted to achieve head pose estimation in academical and industrial fields, but a truly efficient application of achieving low-cost and easy-to-use performance still remains a difficult task. Also, head pose estimation has fewer generic solutions or established systems despite the demand compared to other computer vision techiniques such as face detection and recognition. In this thesis, a monocular vision head pose estimation system is introduced. To increase usability and achieve algorithm efficiency, several aspects of the system have been considered in development, such as a completely automatic process, robustness against noise, identity invariance, and fast computation time. Various approaches for each stage of the system have been analytically tested to select the best methods for each stage.MCT-based AdaBoost (Modified Census Transform-based AdaBoost) is used for facial feature detection. It has been tested robust against noise, identity, and facial expression for automatic and identity-invariant performance. Extracted features are evaluated with a novel projective geometry algorithm. The point-based geometry algorithm based on weak perspective projection combines several advantages of other pose estimation algorithms. Virtual head model is automatically recovered based on the learned data from laser scan database. Unscented Kalman Filter is used for smoothing perturbations generated from estimation errors. Errors or self-occlusions from feature extractions are corrected by a novel nonlinear stochastic system. Experiments of the system on public video databases exhibit real-time performance with lower errors than previous head pose estimation systems. The head pose system presented in this thesis achieves usability and efficiency, and it is projected to be embedded on low performance computing systems such as smart phone and CCTV.