Multi-distribution model for background subtraction in long-term video surveillance system

The basic performance requirement of exploiting background subtraction approach in long-term color video surveillance system was described. To meet the requirement, an algorithm was also proposed. In this approach, by using the multi-distribution models in lightness and chromaticity spaces, the background model was built. The multi-distribution models were then updated using independent mean and covariance updating rate and model replacing rate. When a pixel could be represented by more than one distribution model, the model which has the minimum similarity distance was updated. The approach also involves the detection of the sudden changes in illumination by the feedback of the information of foreground pixels and the suppression of shadow influenced by lightness information. Experiments show that the method satisfies the demands of long-term video surveillance.