The Research And Realization Of Moving Target Detection In Real-time

In many fields, the application of the computer is more and more extensive, and thesurveillance technology with videos also plays a more important part on it, which givenincreasing attention in its intuitive, convenience, security and comfortable with economy. Inmany businesses like bank, the public traffic, the surveillance in public places and etc. thevideo surveillance system is used widely, and the demand to the surveillance system is alsomore intellectualization.The paper analyzes some typical moving object detecting method. Aiming at the demandof surveillance system, we propose a background reduce method based on Gaussian MixtureModels and Histogram Comparison, which introduce the Histogram Comparison into theupdate of Gaussian Mixture Models. The method compares the difference between theneighbor frames on Histogram Comparison, calculates the factor of Histogram changes, andthen judges if there is a large illumination change. Aiming at the situation of largeillumination, the Gaussian Mixture Model updates itself adaptively based on the degree ofillumination change, and contains the original updating method when there is no largeillumination change. The method not only increased the detecting effects in the complexsituation such as large illumination change, but also increased the speed of establishing theGaussian Mixture Model. The performance tests as well as experiment results indicated themethod is accruable and real-time, which could detect the moving target as fast as possibleand satisfy the need of real-time tracks.When using the original background reduce method, it is very easy to fail, which makethe error rates of false-positive and false-negative judgments increasing badly. To avoid this,the paper introduced the Haar-like feature and AdaBoost Cascade classify into the detection ofhead object, which will optimize the classify model with head object. Comparing with otherfeature extraction method, the Haar-like feature extraction method has higher accuracy, whichindicates the method could describe the heat target feature more clearly.At the end, the paper simulates the recent research achievement in the software platform, designing out the software for the purpose of test. The theoretical basis of the software is themixture of the background reduce method based GMM and the pattern recognition methodbased Haar-like feature. The paper runs a group of tests on the software platform version one,and the experiment results prove the forecast which indicate the method could resolve theproblems causing by the large illumination.