Study On Algorithms Of Moving Object Detection In Intelligent Surveillance System

Moving object detection is one of the most important subjects in computer vision, whichcombines advanced technologies in artificial intelligence, image processing, patternrecognition, automatic control and other relative fields. It has been broadly applied in military,medical treatment and intelligent transport and other fields. In this paper, algorithms ofvideo-based moving object detection has been studied quite deeply, which would be veryuseful to be utilized in intelligent surveillance system.In this paper, it introduced the important research value of intelligent surveillance systemin contemporary society at first. Moving object detection plays an important role because it isbase of the entire system. Then this paper summarizes some methods of existing movingobject detection: temporal difference method, background subtraction method and opticalmethod, analyses and compares each method, and indicates their advantages anddisadvantages. At last, some approach after detection is studied, including shadow detectionand removed method, morphological filtering processing method and motion region obtainedmethod.For the difficulty faced in the practical application of moving object detection and themain problems need to solve, a new background modeling algorithm is proposed. Thismethod combines running average and codebook model, which are classic moving objectdetection methods. Both of them belong to background subtract method. Running averagemethod has a fast speed to build background model. However, it has poor performance incomplex environment; Codebook method builds several backgrounds for each pixel, so it hasbeen used to model complex background. Codebook method does have some disadvantages.It consumes much time to determine which codeword will be the best match. This papercombines codebook method with running average method. Codebook is a multi-background model while running average has fast speed. They make up the shortage of each other. It ismore effective than either single method.Codebook of each pixel is made up of several codewords. Every element in codeword hasrespective function for the whole background model. This paper takes away the elementdefined as the longest interval that the codeword has not recurred. It is calculated by othervariables in codeword instead. This improvement on codeword would not change the idea oftraditional codebook. The space of every codeword is reduced. Moreover, the whole spacetoken by model is reduced.Furthermore, there is a training process in the initial establishment phase for codebookmodel. This paper replaces the unified treatment by two–stage training process. In the firsthalf of training process, new codeword is built for input pixel. However, there is no need tocompute whether the existing codebook is redundancy. On the contrary, no new codeword fornew pixel while excess codebook is removed during the later half. The improved codebookmodel is closer to true background.By comparing with related algorithms (running average and traditional codebookmethod), and Mixture of Gaussians (MOG), which is generally considered as a good methodto extract foreground, the new method proposed by this paper is as fast as running average,and the foreground image is as good as MOG. At the same time, it takes less than half space oftraditional codebook. It has apparent advantage in the respect of real-time, memory usage androbustness. It is valuable for intelligent surveillance system.