Bayesian sequential face detection with automatic re-initialization

This paper proposes a probabilistic search algorithm to boost the computational efficiency of face detection in video sequences. The algorithm sequentially predicts the probability distributions of face region parameters in the current video frame given a sequence of past frames, and automatically re initializes the prediction process at scene changes in the given video sequence. A Bayesian criterion is derived as a determinant of likely face regions among the collection of face candidates generated by a sub window-based face classifier. The Bayesian scheme also enable probabilistic marginalization of multiple outputs of the face classifier. Experimental results on a test sequence of 500 frames of broadcast video, containing 450 faces, demonstrate that the proposed approach achieves a detection speed roughly double that of a baseline detector included in the open-source computer vision library OpenCV, without sacrificing detection performance.