Research On Video Object Segmentation Algorithm Based On Multiple Layer Background Modeling

Intelligent video surveillance systems have been widely used in many fields. A typical video surveillance system needs to complete the segmentation, tracking and analysis of moving video object. The video object segmentation is as good as half over. Background modeling is one of powerful tools for target detection. It’s not easy to find a background modeling method that with low complexity, high efficiency and robustness, and can handle all scenes. For this case, the moving video object segmentation method is researched in the paper based on the theory of Mean Shift, video image processing and codebook. This paper is concerned mainly with improving the flexibility, accuracy and so on of the method and mainly completes the rrulti-layer background modeling and object segmentation under the dynamic background.The main research of this paper is as following:(1) With consideration of the problems of background modeling such as accuracy and memory, a multi-layer background modeling based on Mean Shift and codebook method is proposed. First, each position of the pixel gray value is analyzed statistically. That is to say, with the use of Mean Shift algorithm, then carry the pixel gray value through cluster analysis and found the background model by select several clustering centers with larger weight. Second, we use the background model to extract the moving video object. Simultaneously, we should update the background model. Because of using the multi-layer background models, the experiment results demonstrated that the proposed method improved the accuracy of the segmentation, reduced the complexity of the algorithm and reduced the demand of memory.(2) With consideration of the problems of moving video object segmentation under the dynamic background and the purpose of adaptive background updating, a coarse-to-fine algorithm of foreground detection is proposed. In the background modeling stage, we build the multi-layer piece model based on feature vector first, and then build the multi-layer pixel model based on Mean Shift. In the foreground detection stage, we use the multi-layer piece model to filter out most of the background, and then use the pixel model to get the accurate results. The proposed methods, which combine the piece information with the pixel information, have shown the advantage in the speed and the accuracy of the algorithm. The experiment results shows that the proposed method can give a very good result under the static background and can extract a very completed moving object from the dynamic background. The proposed method applied well in the indoor and outdoor surveillance video sequence, single and multiple object segmentation, and the segmentation under the dynamic background. This method can make an accurate result and can satisfy the demand of the video surveillance system.