Horizontal Face Pose Estimation using Dynamic Bayesian Networks

S.A. Suandi (Malaysia), S. Enokida, and T. Ejima (Japan)


Computer Vision, Face Pose Estimation, Dynamic Bayesian Networks, Eyes and Mouth Tracking


This paper describes a method to estimate horizontal face pose from the detected facial parts using dynamic Bayesian networks (DBNs). Facial parts, which are eyes, pupils and mouth, are fully utilized to (1) construct a model known as ”head cylindrical model (HCM)”, and (2) compute three observation values for inference using DBNs. HCM is de signed originally from some anthropometric data of aver age pupils distance and head width. This model provides the head center, which behaves like a shaft or a joint where rotation of the head is made about this joint (similar to hu man head biologically). Using this model, we simplify the difficulties to compute center of human head and reduce sensitivity to noise while tracking, especially when head moves or pans. Once HCM is constructed, it will be fit into the head and kept inside it and the head will always be kept inside it along the tracking process. A scheme to maintain its position relative to face motion is also explained. The contribution from this work is that our proposed method fully utilizes detected features from tracking module and does not require any additional pre-processing stage, thus, it is computationally cheap and suitable for real-time appli cations. Results also reveal that our method achieves good results and performs at real-time.

Important Links:

Go Back