The Estimation And Implementation On Crowd Density Estimation In Intelligent Video Surveillance System

Intelligent video surveillance is one of research hotspots in the fields of computer vision. As thenumber of customers increases in the city, how to monitor and control the crowd in the publicbecomes very important. Crowd detection and crowd density estimation in the scene not only help thedepartments to inspect unsafe situations in public concourse in time, but also enhance supervisingperformance of system. It would have a great contribution on maintenance of safe order in publicconcourse.Current approach and method of crowd density estimation are studied in this paper. Theimproved Gaussian Mixture Model(GMM) is proposed to establish the background model for low andmiddle density crowd image. Unlike the conventional method, according to the characristic of crowdimage, the feature of background model can be described better by the chosen mean and the mean ofdeviation of images. The interference pixels of moving target for background modeling were thensolved by pixel filtering we developed, and the crowd foreground image is finally revised accordingto the difference of S (saturation) component between original image and background image. Theexperiment results have shown that proposed method can reconstruct the background model quicklyand it can improve the accuracy of crowd density estimation for middle or low density crowd image.The Support Vector Machine(SVM) is applied for estimating middle or high density crowd. First,image textures are extracted via gray-level co-occurrence matrix and textures in different directionsare selected after correlation analysis. The improved SVM is then used to classify the crowd image,whose parameters are rapidly selected on the basis of segmented dichotomy. The experiment resultshave shown that the textures in different directions decide the accuracy of estimation of crowd imagesunder different scene. Moreover, the generation performance of SVM is better than BP neural networkclassifier, and it can enhance the efficiency of crowd density estimation. At the last, according to thecharacteristics of intelligent video surveillance system, multichannel crowd density estimation systemis implemented by multithread processing based on single crowd density test system, and it canimprove the ability of the existing intelligent video surveillance system.In the paper, the crowd density estimation proposed has realized in intelligent video surveillancesystem, and it can meet the accuracy and real-time request in vedio surveillance system. The proposedmethod has the broad application prospect in public surveillance system, such as airport, subway andpublic squares.