Background Modeling from Surveillance Video via Transformed L_1 Function

摘要

Background modeling from surveillance video plays a key role in event detection and human action recognition. Its goal is finding moving objects in video that are independent of the background scene. Among many background modeling algorithms, robust principle component analysis is a recently popular technique, which characterizes the moving objects via L_1 norm. However, L_1 norm often leads to inaccurate solution. To overcome this limiting, this paper proposes a background modeling method based on transformed L_1 function (BM-TL1). The motivation is that transformed L_1 function can interpolates L_0 and L_1 norms by tuning a parameter. Another merit of the transformed L_1 function is it enjoys closed form iterative thresholding function, and thus can be optimized efficiently. Experiments demonstrate the effectiveness of the proposed methods.