Research On Crowd Density Estimation And Abnormal Event Detection In Video Surveillance

With the development of modern society,the number of the world's population is increasing rapidly.In recent years,many accidents have taken place in densely populated public places,and more individual extreme behaviors have brought great harm to people's life.How to discover and deal with potential safety hazards in the crowded places has become an important research subject.According to the differences of the characteristics of the population,there are two kinds of methods: methods based on pixel statistics and methods based on texture features.In the case of high density population,especially with occlusion of people,method based on pixel statistic has a low detection rate.For methods based on texture features,time complexity is large,and the monitoring area is not always maintained at a high density.When the population density changes,the texture feature of the population is not obvious.In this paper,I have researched the method of population density estimation based on pixel statistics.First,I divide the dataset into low density,medium density,and high density,then I extract the foreground,achieve perspective correction,build the model,and detect abnormal events using foreground image.For detection of abnormal events in video surveillance,there are two methods: method based on individual target trajectory and method based on local low level feature representation.For abnormal event detection,method based on individual target trajectory,single objective detection and tracking in video is first taken,and the moving trajectory of the target is matched with defined anomaly event model.The method relies on the moving trajectory,but the trajectory tracking technology is still not mature enough to provide reliable and accurate motion trajectory information.Besides,the variety of abnormal events is large,the definition is fuzzy,and it is difficult to define normal events and abnormal ones with the moving trajectory of the target.For methods based on local low level feature representation,low level features are used to represent the normal and abnormal events in the monitoring area.The method is suitable in crowded scenes.However,description using edge and corner features has low accuracy.In this paper,a new method for detecting abnormal events is proposed,by extracting the foreground of the image and feature extraction with the feature template,and then the feature data is input to the model to detect pedestrian abnormal events.In this paper,experimental environment is Windows8 system,AMD A8-4500 M APU,4G memory,Visual Studio 2013,OPENCV platform,datasets with PETS2009,UMN,UCSD,BEHAVE Interactions Test Case Scenarios.