A nonlinear background updating scheme

Background subtraction (BS) is one of the key techniques in video surveillance applications. To cope with the variations occurred in a scene, including illumination change, the background (BG) model should be updated timely to adapt to those changes. Generally, the BG model is updated by a linear moving average scheme, but this will produce “ghost” trails in the evolved BG model since foreground objects may be merged into the BG model during the updating. In this paper, we propose a novel nonlinear BG updating scheme: the updating rate of the BG model is modified by the absolute difference between the current input and the BG model at the previous moment, and this can suppress the artificial “ghost” trails behind the moving objects in the updated BG model. In order to study the effect of the nonlinear BG updating scheme on the detection results and the evolved BG model over time, we integrate it into two representative BS algorithms: the frame difference and the ViBe algorithm. The nonlinear BG updating scheme can not only extend the discrete updating states in the ViBe algorithm to a continuous one, but also implicitly provide a multi-temporal scale for complex scene. We evaluate the proposed nonlinear BG updating scheme on a database which consists of over 60,000 frames in 21 videos representing 4 scenario categories, and the experimental results demonstrate its efficiency.