Background modeling based on spatio-temporal patch

A novel spatio-temporal patch based background modeling (STPBM) approach was proposed. Spatio-temporal patches (bricks) are utilized to characterize both the appearance and motion information of objects in videos. It was observed that, under all possible illumination conditions, all the bricks at a given background position lie in a low dimensional background subspace. In contrast, bricks with moving foreground are uniformly distributed in original space. Then an efficient online subspace learning method for capturing the background subspace was presented, and the incoming bricks with moving foreground could be detected according to their distance to the background subspace. Experimental results demonstrate that, compared with traditional pixel-wise or block-wise methods, our approach is more insensitive to drastic illumination changes and capable of detecting dim foreground objects under low contrast.