The Research Of Crowd Density Estimation And Crowd Statistics In Video Surveillance

Along with the growth of world population and the rapid development of socialeconomic, mass incidents are increasingly frequent. Therefore, crowd safety issue hasbecome a hot issue. The rapid development of video surveillance technology is animportant technical support for controlling group incidents. Crowd density or theexact number of people is critical basis of crowd management.Traditional visual surveillance applications are almost manual. This methodrequires a huge amount of work to monitor and the attention of surveillance personnelwill distract with the growth of the monitoring time, it is likely to cause big delay andmistakes. And afterwards, even found the problem through searching and analyzingstored data, it is too late. In recent years, large-scale of crowd density estimation andpeople counting algorithms are emerging out. But the existing crowd densityestimation and crowd statistics algorithms have many disadvantages, such as real-timeis not very good, the accuracy has more space to be improved and so on. This thesissummarizes and compares the mainstream algorithms of crowd density estimation andcrowd statistics. And based on these algorithms, we propose a new algorithm.Firstly, this paper comprehensively utilizes the advantage of feature pointstatistical and Grey Level Dependence Matrix, and combine support vector machineregression methods proposed a new crowd density estimation system. As the featurepoint statistical performs well in low density crowd and the texture of Grey LevelDependence Matrix is suitable for high density crowd, the proposed system has arelatively high accuracy in the whole density range. Without extract foreground imagebased on feature point method, so this density estimation system has well real-time.Compared this method with some classical crowd density estimation methods, we canconcluded that the proposed method has higher accuracy.Finally, the author uses primary component analysis to analyze correlativerelation between the texture indicators and presents a crowd statistics method basedon three dimensional texture statistical analysis method to estimate the number ofcrowd. Through the method of primary component analysis, we can estimate thenumber of crowd by using first primary component as one dimensional statistic andfind the most important indicators are contrast of each direction. With the comparisonexperiments of one dimensional statistic, three dimensional texture statistics andtwelve dimensional texture statistics, the author find out the most appropriate crowd statistics method is the method by three dimensional texture statistical analysis.Experimental results show that the proposed method of crowd statistics andcrowd density estimation can get a high accuracy and meet the real-timerequirements.