Background Modeling In Nonstationary Scenes For Object Detection

Intelligent system for video surveillance is one of the hottest fields in digital video research. It integrates isolated video surveillance products into a single surveillance system via advanced technologies such as digital signal processing, pattern recognition, artificial intelligence and automation so as to intelligentize the system. The key techniques for this kind of system are object detection, recognition and tracking, among which object detection is preliminary step for those two others.Background modeling is always an important issue in accurate detection of moving objects. In this paper, we present a novel non-parametric foreground-background model which explores the complex temporal and spacial dependencies in nonstationary scenes. The model adapts to scenes which contain small motions such as tree branches and water ripple. The model estimates the probability of observing pixels'5-dimensions feature vector which represents its intensity values and spacial position information. The model is built and rolling-updated by kernel density estimation (KDE). Finally, we use a maximum a posteriori-Markov random field (MAP-MRF) decision framework to segment the foreground and background by solving a graph-cut. Extensive experiments with nonstationary scenes demonstrate the utility and performance of the proposed approach.