Real-time dynamic background segmentation based on a statistical approach

Background modeling is usually the first step in vision-based surveillance systems. Subsequent foreground segmentation can then be performed by comparing the variations between the current image and the reference background of the monitored scene. Various approaches have been proposed to deal with this issue. They differ in the type of background models used. However, these approaches emphasize only what the distribution of the background looks like, not what the real actions of the background are taken place during some period of time. In this paper, we propose a real-time background model that can automatically self-adjust to the scene changes. The experimental results show that the proposed background model has better performance over others in terms of noise suppression and the preservation of foreground details. In addition, our model can also operate correctly at night. Furthermore, it can effectively resist shaking of cameras and objects.