Background subtraction using spatiotemporal condition information

Abstract

To improve background subtraction performance in dynamic scenes, a novel background subtraction method using spatiotemporal condition information (SCI) is proposed. To obtain SCI, a spatiotemporal neighborhood was constructed based on the center surround visual saliency model. To reduce the errors in the background subtraction result, a neighborhood weighted spatiotemporal condition information (NWSCI) was utilized for pixel classification by considering the similarity of neighborhood pixels. A joint cascade and hierarchical framework was introduced to reduce computational cost by rejecting the unchanged regions in videos before background subtraction with NWSCI. Experimental results show that the proposed method can effectively detect moving object in dynamic scenes in real time, and performs better compared to the existing methods with lower resource cost.