MRF-based adaptive approach for foreground segmentation under sudden illumination change

Background modeling is an essential processing component for many video applications. However, most reported background modeling algorithms, which are based on the intensity difference detection, fail to handle sudden illumination change in monitored scenario. In this paper, a novel Markov random field (MRF) based probabilistic approach for background modeling in video surveillance applications is presented. The proposed framework takes both intensity and texture observation into account and fuses them spatially and temporally in an adaptive way to cope with sudden change in illumination. Using Gibbs sampling to solve the MRF in a maximum a posterior framework, proposed algorithm achieves real-time performance. Both visual and quantitative experiments in several sequences of indoor scene demonstrate the effectiveness of our algorithm.