Research On Crowd Anomaly Detection Based On Gaussian Mixture Model

With the popularity of surveillance system,using intelligent video surveillance system to replace the traditional artificial supervision has become a trend.Especially,crowd anomaly detection has become a research hotspot of intelligent video surveillance.Crowd anomaly detection is mainly aimed at scenarios where there are large flows of people,such as railway station,airport,campus,etc.Abnormal events in the scene should be automatically detected,with alarm in time.We can minimize the damage caused by abnormal events in crowds.It plays an important role in the protection of personal safety.However,the crowd density is sometimes higher in the real scenes and anomaly has various types,so how to use the computer vision technology to analyze the crowd scenes has become the development trend in recent years.In this paper a method based on Gaussian mixture model(GMM)is proposed for crowd anomaly detection,the research work and innovation of this paper are as follows:In the preprocessing stage,we get the moving region of the frame in the video.That is,the region of interest(ROI).Then fill these holes of ROI with morphological method.We first achieve the final ROI which eliminates the interference of background and get the completely moving target.Then extract the motion information of video by Pyramid Lucas-Kanade optical flow and SIFT(Scale-invariant feature transform)algorithm.The crowd is analyzed as a whole to omit some complicated processes,such as pedestrian detection,target tracking,etc.It still has a good effect on those scenes of illumination fluctuations and crowd occlusions.The abnormal events are detected by dividing the image into non-overlapping blocks.Because the motion features of different regions are different,different blocks are trained with GMM on the data of normal scenes.It replaces the artificial experience judgment.Each block is detected whether there is an abnormal event with corresponding trained GMM.This method improves the accuracy of crowd anomaly detection and also locates the position of the abnormal events.In this paper,the simulation experiments are carried out on two public abnormal events database,UMN and UCSD.The proposed approach is compared with the state-of-the-art methods and we use ROC and AUC as evaluation standards.The experimental results show that the proposed method can effectively detect the abnormal events of different scenes and can locate the positions of abnormal events.The performance is better than other previous methods in the same evaluation standard.