Motion Detection Algorithm Based On Background Modeling In Surveillance Video

With the increasing progress and development of modern science and technology,surveillance cameras have quick access to all areas of society. These cameras arerecording a lot of surveillance videos which are few specialized to check up. Only whensome serious incidents happen, people use the videos as evidence, but they often do notreflect their proper values. Therefore, people begin to pay more and more attention tosurveillance video processing.In the area of surveillance video, the most important task is to make judgments byanalyzing the movement of people and objects’ behavior, so as to prevent crime. Thebehavior analysis of moving objects in the video is based on the technology of motiondetection and tracking.Therefore, this thesis will focus on detecting the foreground objects in surveillancevideo, by observing the distribution of background pixel values in RGB space,formulating hypotheses, proposing background pixel value distribution model, and thenverifying the effectiveness and accuracy of the model.However, most background modeling algorithms have inaccurate depiction problemof the background pixel value variation. Their assumptions of background pixelvariation are not in line with the actual situation. For example, the mixture of Gaussiansalgorithm assumes that the three components of RGB space are mutually independent.In order to solve the above-mentioned problems this thesis mainly does the followingwork:①Study the various methods in the literature used for moving objects detection,and draw a roadmap for moving object detection by sorting out these literature.Research the famous moving object detection algorithms, mixture of Gaussiansand codebook, and give out a detailed analysis of the two algorithms’descriptive ability in background modeling area, and analyze the reasons ofinaccurate assumptions. In the end, verify the pixel value distribution in theRGB space on the actual video sequence by sampling verification experiments;②Propose a novel random sample consensus based regression backgroundmodeling algorithm through the sampling experiments. The algorithm modelsthe characteristics and laws of background pixel value’s change as a linearregression model. It changes the RGB three-dimensional problem into one-dimensional problem;③Propose a principal component analysis based codebook background modelingalgorithm through the sampling experiments. The algorithm models thecharacteristics and laws of the background pixel value’s change into an RGBspace cone. And evaluate the central axis of the cone through principalcomponent analysis algorithm;④Make a comparison of the proposed algorithms with mixture of Gaussians andcodebook algorithm through the actual detection experiments and ROC curveanalysis.