Hierarchical CodeBook for background subtraction in MRF

Foreground detection is the key low-level fundamental work in intelligent video surveillance. This paper proposed a hierarchical background subtraction algorithm consisted of block-based stage and pixel-based stage for it. In block-based stage, obvious backgrounds got detected via block-based CodeBook, leaving spatial relations among suspicious foreground pixels undestroyed. Pixel-based stage further eliminated the left background pixels with the introduction of spatial and temporal relations in a MRF-MAP framework. Then comparative experiments were conducted to evaluate the performance of the scheme in three dimensions – detection accuracy, update speed and memory consumption. Proposed approach possesses the highest detection precision and consumes the second least memories. And the update speed is of real-time level.