Efficient and accurate background subtraction using multilevel background model

Background subtraction is an essential step of intelligent video surveillance and has got a lot of interest among researcher community in recent years. It has a critical impact on the performance of object tracking and activity analysis. In this paper we propose a new multi-level background modeling to overcome dynamic background problem. At rst, an image is segmented to foreground or background in larger window(12 12) level and results are rened by smaller windows(6 6 and3 3)level and pixel based operation. Our pixel based segmentation section uses VIBE method , which is fast and needs less memory to conserve background model components, with some modication to detect shadows. Subtraction is done in coarse levels rstly, and resulted foreground are investigated more by smaller windows(ne levels). This makes the algorithm to be more ecient. A new once-o background changing detection and model updating is proposed to make our algorithm as accurate as possible. The last part of our algorithm is enhancement where, we have used morphological operators in order to improve the subtraction quality. The approach provides us with many advantages compared to the state-of-the-art. Experimental results clearly justify our strategy.