Background Subtraction Algorithm Research

Background subtraction is a key technology of computer vision and target tracking. This thesis is based on the current development of background subtraction. According to the problems and challenges in research work nowadays, it tries to improve the real-time performance and veracity of background subtraction, and at the same time, to resolve the problem of varying illumination and dynamic textures. Some background subtraction algorithms are approached and a lot of experiments are done to validate their feasibility.First, considering the static background, an improved background subtraction algorithm is approached. The algorithm is based on three histograms of RGB. It not only improves the veracity of background subtraction in gray video but also decreases the computational capacity of the algorithm based on code-book construction. Three histograms are constructed according to the RGB data of an image sample sequence from a color video. The RGB value which has the highest frequency in their histograms is found to construct background model. Then the color distortion and brightness range are analyzed for per pixel in the current image and background model. Background is subtracted according to both color distortion and brightness range. This algorithm improves the real-time performance and veracity of color video background subtraction.Secondly, a new algorithm is approached because of the influence of varying illumination on the performance of background subtraction algorithm. The image is carved into sub areas before analyzing vector linear dependence of each pixel. The computational capacity of analyzing vector linear dependence is very huge, and the image in video sequence can be carved into three kinds of sub areas which are no variant area, varying illumination area and varying content area, in which the first two kind areas have no targets. If the image is carved into three kinds of sub areas, not only the targets can be found just by analyzing the pixel's vector linear dependence in the third kind area but also the computational capacity is largely decreased. The algorithm based on analyzing vector linear dependence of each pixel in sub area improves the real-time performance to a huge extent.Finally, the video with dynamic texture in background is analyzed. The dynamic area is found and a model in state space is constructed. The next state of dynamic area is forecasted by Kalman filter. The target is found by analyzing the difference between the current image and the image forecasted by last frame. The algorithm improves the ability to subtract dynamic background and develops the algorithms which are used nowadays. It is of certain reference value.The realization work of all the algorithms in this thesis is done with MATLAB .It illustrates that the algorithms are feasible and of application value.