Face Tracking via Block Texture Feature Based Mean Shift

Face tracking plays an important role in many computer vision applications such as human-robot interaction and visual surveillance. However, it is still a challenging problem, due to various factors related to illumination, cluttered background and poses variations. In this paper, we introduce a novel feature descriptor, namely block binary pattern (BBP), to represent the appearance model of face for the tracking tasks. Compared to local binary pattern (LBP), BBP has the advantage of capturing multi-scale structure, while preserving the robustness to illumination and appearance variations, and meantime, it can be extracted in realtime for real-world applications. Based on the BBP features, we use AdaBoost strategy to select a discriminative features pool. These features can be considered as the prior appearance model of face. We use similarity-based mean-shift, which is the extension of original mean-shift, as the face tracker. Experimental results on challenging sequences validate the effectiveness of our method for face tracking.