Research On Moving Foreground Segmentation Technique In Videos Surveillance

As the society enters IT era, video surveillance system plays an important role in many fields. Meanwhile, the video processing technology has also become an important work of scientific topics. As a fundamental task of video processing, video foreground segmentation has a great impact on the subsequent steps. At present, existing video segmentation algorithms still have some lacks in aspects of accuracy and robustness. Therefore, how to develop one kind of foreground segmentation algorithm with high precision and robustness is of great meaning to both research and practice.In this thesis, at first, a detailed summary and analysis of the domestic and international research results in the field of video processing has been carried out. This work mainly focuses on background modeling and moving foreground extraction in video segmentation processing. Through the analysis of the basic theory of several popular algorithms and the comparison among algorithms, we focus on video segmentation algorithms basing on codebook background model. At last, an improved method is proposed to overcome deficiencies of existing methods. The main research contents and results are as follows:Firstly, to overcome the deficiency of codebook cutting strategy and to speed up codeword matching process, an optimization method is proposed. And, to relieve the impact of the matching efficiency brought by the disorder of the original codewords, a sorting algorithm is developed. Based on the original time filtering, frequency filtering is introduced when the brief codebook is generated. Experiments show that the modified algorithm is more efficient without deteriorate the segmentation results.Secondly, in view of moving shadow interference problem in video segmentation processing, an improved method is proposed. First, as moving shadow and foreground targets show the different nature in the color space model, the shadow could be preliminary filtering according to color similarity. Then, the target location is further determined through a gradient threshold segmentation method. At last, the foreground targets are obtained by combining with the initial partition generated from background subtraction process. Experiments show that the improved algorithm can effectively eliminate moving shadow and can inhibit the residual noise outside the target areas. Finally, based on the research results of this thesis, a video foreground segmentation system is developed. Experimental comparisons have verified that our improved algorithm outperforms common methods.