Research Of Moving Object Detection Algorithm Based On Background Modeling

Moving objects detection is an extremely significant part of the research in computer vision, and analyzes the information of moving objects by means of image processing, which is smart and plays a very important role in many vision applications such as smart video surveillance, intelligent transportation, human-computer interaction, activity recognition, object tracking, behavior understanding, precise guidance etc. At present, there are three main categories of the algorithms of moving objects detection based on optical flow, frame difference and background modeling respectively. The method based on background modeling takes three important steps of building the background model, classing pixels and updating the background model to detect moving objects, in addition, it has the advantages of more modeling methods and eigenvalues and adapting to the change of the scene with the background model updating, all of which make the method based on background modeling become one of the important domains of the research in moving objects detection.The algorithm of ViBe(visual background extractor) which makes use of the relation of neighborhood to build and update the model, and therefore the relation of space can be effectively used, which is of wide applicability, fast modeling speed, and good robustness. But the method has some limitations in application, such as the performance decline in the dynamic scenes in which there are many rock spots or there are moving objects in the initialization frames. This thesis focuses on the algorithm of ViBe to improve the applicability in the special and complex scenes. The main work in this thesis is as follows:(1) We present an improved method called multi-frame initialization modified ViBe. It builds the background model through the mid-value of the concentration region based on statistical characteristic, and modifies the strategy of updating background model. The improved method solves the problem that the method of ViBe will get a lot of false detection when there are shaking points caused by branch or surface of water in the scene. We take the test on scores of dynamic background databases such as canoe, fall and so on. The results indicate that the algorithm improved can effectively reduce false detection and improve the detection.(2) In this thesis we present a supervised detection algorithm called SViBe, which introduces the check of the result of the classification of pixel, helps to solve the problems of pseudo moving object caused by the moving object in the background initial frame image and sudden movement of the stationary objects and the moving object becoming stationary if uses the algorithm of ViBe, which speeds up the pseudo moving objects fusing into the background. The algorithm of SViBe has been used in the tested on many databases of pedestrians, office, and all the results shows that SViBe can be more effective in the special scene represented above.