A Robust Background Initialization Method Based on Stable Image Patches

Abstract

The estimation of the stationary background from the video sequence is a challenging problem in Computer Vision. The process can be easily affected by illumination changes in the video sequence and estimate inaccurate background image. To solve this problem, this paper improved the Labgen, the best performance method in the SBMC2016, and proposed a robust background initialization method based on the stable image patches. This method can generate accurate background image when the given video sequence has strong illumination changes, and this method also has strong robustness. First, the most stable illumination condition sequence is selected for the next background modeling. Then each frame of video is divided into patches of the appropriate size. Select the most stable image patches by frame-by-frame comparison over the entire video sequence; and use the temporal median filter for the subset of selected patches to generate background image. The experimental results on the SBMnet dataset show that the proposed method can generate accurate background image, and perform better compared with the latest method when the video sequence has strong illumination changes.