Moving Objects Detection Method Based On Gaussian Mixture Model

With the development of computer science and communication technology, the areaof intelligent security is also developing rapidly, as a result, people’s awareness ofsecurity is also rising. The public places such as banks, shopping mails, transportationhubs and high-grade community are equipped with intelligent monitoring system, as thecore technology of this system, the moving target detection has received wide attentionfrom researchers at home and abroad. The moving target detection is not only theintersection of fields of image processing, pattern recognition and computer vision,butalso the premise for locating target tracking, motion analysis and other techniques, sothe research on it has great significance.There are mainly three moving target detection methods: optical flow field,inter-framedifferencemethod and background-difference method. The first method isproposed by Horn and Schunck, on the one hand, it brings in the optical flow constraintequation creatively, on the other hand, it figures up the motion vector field of the sceneaccording to the grey information contained by the image pixel and makes a distinctionbetween the foreground and the background through the difference on the motion vectorbetween moving targets and backgrounds. However, because of the large amount ofcalculation and the bad effect on real-time, it has not been widely applied. The biggestadvantage of the second method is simple and fast, through which the subtraction isdone on the neighbour video frames, resulting in good real-time performance, but formoving slowly or stationary objects in the scene, it may extracts a hole or even nothing.The most widely applied technique in moving objects detection is the third method. Thetechnology of applying Gaussian Mixture Model to construct the background proposedby Stauffer and Grimson has been widely recognized, which can robustly overcome theproblem of slow change of the light and slight shake of branches and so on. However,when the background illumination change d suddenly, it is difficult to detect the foreground. Moreover, it has an disadvantage of large amount of computation and poorreal-time performance.Considering the shortcomings of Gauss Mixture Model, we did some researcheswhich are mainly as follows:First，we propose every other pixel modeling, which may greatly reduce the burdenof large computation of GMM. If the pixel P(i, j) is detected as background orforeground, the pixel at P(i+1, j), P(i, j+1) and P(i+1, j+1) may be considered asbackground or object too; Second, the paper considers that all model of GaussianMixture Model in the modeling process may be background model, while the noisepoints have been abandoned probably when matching. So all the model matched areadded in accordance with their weights to build the background model; Last, when thereis a sudden illumination change in background, GMM will no longer be accurate, for itwill detect the most area of an image as foreground. Aiming at the problem of GaussianMixture Model, a new Gaussian Mixture Model with adaptive lighting is proposed. Sothe influence on the model by the illumination mutation has been weakened, as a result,the performance of the mixed GMM has been improved.