A Bayesian framework for robust human detection and occlusion handling human shape model

One challenging aspect of automated surveillance for real environments is the occurrences of various difficult scenarios brought about by practical unconstrained settings. We address foreground detection for automated surveillance under the following challenging situations: i) foregrounds being partially hidden due to close similarities to the background, and ii) foregrounds representing multiple objects being inseparable, forming a large contiguous blob due to occlusion. To build a robust system, we present a new foreground detection framework based on Bayesian formulation, comprising both bottom-up and top-down approaches. We first propose a region-based background subtraction and a localized spatial segmentation scheme as the bottom-up steps for foreground detection. We then incorporate a human shape model as the top-down step for foreground validation and occlusion handling. Segmentation is obtained when a maximum posteriori value is found, corresponding to the best description about foregrounds given by the approach. Such integration of bottom-up and top-down approaches leads directly to more robust performance in handling challenging situations within hostile real environments. Promising results are obtained when the algorithm is tested on real video sequences captured from a live surveillance system that operates at a public outdoor swimming pool.