Improved Visual Background Extractor Based on Motion Saliency
Zhihu Wang, Xiaoqing Shen, Jian Sun, Bin Qiu, Qinghua Yu · 2020
Motion detection has been widely applied as the fundamental step of video analysis. In this paper, we present a technique for extracting the moving objects from video sequences, which is based on the off-the-shelf visual background extractor (ViBe) method. In order to improve the performance of ViBe, a series of modifications are introduced. First, inspired by the saliency detection model, the differences between the pixels of the coming frame and all its corresponding background samples are accumulated and the motion saliency map is generated, by which it can suppress the inaccuracy of the background model in that it takes all the samples into consideration instead of parts of them. Second, based on the above saliency maps, the modified fuzzy growing method is applied to extract the foreground mask. Experiments on a variety of visual surveillance databases validate the improvements of the proposed modifications. Especially in a scenario where the moving objects stop for a short while, our method can greatly suppress the diffusion of false negatives.