Research On Moving Object Detection In Complex Scenes

Moving object detection is an important research topic in computer vision as well as a fundamental part of many vision applications such as intelligent video surveillance systems, visual navigation, video compression coding, etc. As a research topic with high application value, moving object detection has received increasing attention from researchers in recent years. Plenty of new methods have been proposed and applied to vision applications. However, present moving object detection algorithms are limited to some specific application scenarios. Effectively detecting moving objects in complex scenes is still a very challenging problem, and hence has broad application prospects.The background of the videos that recorded in complex scene is under continuous changing, which brings great difficulties to moving object segmentation. Either the problems of illumination changing/leaf shaking in static background or camera moving in dynamic background, they are key problems to be solved in moving object detection. In this thesis, the research on moving object detection algorithms under both static background and dynamic background in complex scene has been conducted. The main work is as follows:(1)Firstly, a variety of common moving object detection algorithms are respectively studied and summarized. The theoretical analysis and experimental results of these algorithms are given. Besides, a diversity of advantages and disadvantages of all the algorithms are listed and the hot improving spots of these algorithms are discussed.(2) After a comparative analysis of various moving object detection algorithms for static background, aiming at Gaussian background modeling method, a novel algorithm based on single Gaussian model for video background modeling and moving object detection is proposed. This new method divides the video images into several blocks and uniformly models the pixels in the blocks to replace the traditional per-pixel background modeling method. Since the proposed block modeling method matches the characteristics of Gaussian distribution better, it would be more conducive to exploit the advantages of single Gaussian modeling method. Hence, it enhances the capability to deal with complex background and effectively reduces the complexity in operation.(3) With respects to various moving target detection algorithms under dynamic background, this thesis mainly studies the global motion estimation method which is based on moving compensation. Theoretical analysis and experimental verification are made for two different global motion estimation methods-whole pixel method and motion vector method. Simultaneously, the impact of singular motion vector for motion vector based Global Motion Estimation method is discussed when the experimental results are being analyzed.(4) Aiming at the problem that singular motion vectors may affect the effect of global motion estimation method, a novel algorithm based on corner detection and K-means method is proposed in this thesis. Firstly, the novel algorithm selects the region with abundant details in the video frame using corner detection algorithm instead of blocking the images in the traditional algorithm. Thus, it can reduce the generating of singular motion vectors by lowering the misalignment rate of block-matching algorithm. Then the novel algorithm excludes the motion vector generated by the moving target using K-means method. Experiment results show that the novel algorithm can eliminate the influence brought by the singular motion vector in a certain degree.