Mean shift face tracking with dynamic target model update using Bayesian skin classifier

Mean shift based face tracking is able to track face nicely under controlled conditions. It usually fails when there is noise and rapid illumination variations over face. The main reason for failure is that mean shift employs fixed histogram based target model representation. In this paper we present a novel technique of continuously updating the target model histogram for more robust mean shift based face tracking using a Bayesian skin classifier. We used Bayesian skin classifier to learn new skin color features from face templates in successive frames. The classifier use likelihood ratio for extracting new skin color features from each new face template. The Bayesian skin classifier had been used mainly for detection of skin parts in images but we extend this idea to continuously update the target model histogram with learned features to accommodate the dynamics of pose change and illumination in a video sequence for more robust face tracking.