Research On Crowd Abnormal Behavior Detection Algorithm In Intelligent Video Surveillance System

In recent years,with the development of the economy,crowd gathering phenomenon is more and more frequent in some public places.Some uncontrollable mass incidents are likely to occur when the crowd density is too high.Therefore,monitoring the population for high population density and other abnormal events is critical.This thesis studies two aspects of crowd density anomaly and abnormal crowd behavior.The main research work is as follows:1)The general crowd density estimation algorithm is affected by crowd occlusion,so it is difficult to accurately estimate the crowd of different density levels.The thesis estimates the number of crowd based on the foreground pixels.Firstly,the background model of the scene is built,and the image is segmented to reduce the influence of the perspective effect.Then the number of the foreground pixels is calculated,and the thesis uses different proportional coefficients for different positions.A new calculation method of occlusion coefficient is proposed for the problem of occlusion.Then,a crowd density estimation algorithm which only detects the texture feature of foreground image is adopted for the influence of background.2)It is difficult to detect crowd abnormal events,because the anomaly itself is very complex and there are many kinds of abnormal events,so it is difficult to detect all kinds of abnormal events.so this thesis takes the samples containing normal population as the training set and uses dictionary learning to obtain an appropriate dictionary.The dictionary is used for sparse representation of the sample to be tested,and the test samples are classified according to the sparse reconstruction energy value or the residual error.The reconstruction energy and residual of the abnormal samples are higher than that of the normal samples.Since the dictionary of general detection method is fixed,it is difficult to cope with the new situation brought by the increase of video.Therefore,this thesis proposes two online dictionary updating algorithms.The first is to set a weight coefficient based on the usage frequency of the dictionary atom,and replace the atom when the weight of the atom is lower than the threshold value.The second is to manually intervene to update the dictionary when the algorithm detection is wrong.As the dictionary is updated step by step,the ability to distinguish abnormal events is improved.