Abstract: Moving object detection in dynamic background is a very challenging fundamental problem in video surveillance. This paper presents a robust moving object detection method. First, we develop an effective ViBe algorithm against dynamic background by incorporating the scene prior information that is predefined in initial frame. Then, the global GMM models of foreground objects and background are estimated by foreground and background pixels detected by the improved ViBe algorithm. These GMM models are employed to classify every pixel effectively and remove some of the false results. For further alleviating the effects of noises, the superpixel-based refinement is adopted to obtain the final results. The experimental results on the collected video sequence with strongly dynamic background suggest that the method significantly outperforms other moving object detection methods.


Key words: dynamic background, scene semantic prior, ViBe algorithm, appearance consistency, GMM model