A Bayesian Approach to Background Modeling

Oncel Tuzel, Fatih Murat Porikli, Peter Meer

Abstract: Learning background statistics is an essential task for several visual surveillance applications such as incident detection and traf.c management. In this paper, we propose a new method for modeling background statistics of a dynamic scene. Each pixel is represented with layers of Gaussian distributions. Using recursive Bayesian learning, we estimate the probability distribution of mean and covariance of each Gaussian. The proposed algorithm preserves the multimodality of the background and estimates the number of necessary layers for representing each pixel. We compare our results with the Gaussian mixture background model. Experiments conducted on synthetic and video data demonstrate the superior performance of the proposed approach.