Robust compositional method for background subtraction
Xiaochun Liu, Tao Zhong, Dan Fu · 2012
We present a novel compositional method for background subtraction. The presented method can better handle multi-modal appearance of the background in a dynamic environment. In this method, an illumination compensation method based on Local Binary Pattern is used to compensate sudden global and/or local illumination changes, and it is also used for filtering the pixels from background rather than foreground for background subtraction algorithms. Using this scheme, we can improve the accuracy of background model and make our method better handle the high-frequency illumination changes. An efficient compositional scheme is proposed to combine the parameter-based and the samples-based algorithm for background subtraction. Thanks to this scheme, we can get more reliable results than any of them. A new initialization technique for Gaussian Mixture Model based on spatial information is also presented, which can help us get meaningful results from the second frame. Experimental results clearly justify our method.