Robust online video background reconstruction using optical flow and pixel intensity distribution

Obtaining a dynamically updated background reference image is an important and challenging task for video applications using background subtraction. This paper proposes a novel algorithm for online video background reconstruction. Firstly, multiple candidates of background values at each pixel are obtained by locating subintervals of stable intensity in a processing period. Then criteria based on pixel intensity distribution and local optical flows are employed to decide the most likely candidate to represent the background. For the methods utilizing the distributions of intensity values, the decision of determining the background value at a pixel position is based on the observation that the appearance time and subperiod frequency of the background is higher than non-background. An enhanced method using neighborhood optical flow information is adopted for more precise decision with slightly additional computation by identifying the events of covering and revealing of a pixel position. The experimental results show that the proposed algorithm outperforms existing adaptive mixture Gaussian background model and provides robust, efficient background image reconstruction in complex and busy environment.