Video Anomaly Event Detection Based On Deep Learning

In recent years,with the improvement of people's living standard and national quality,as well as the rapid development of digital information technology,all sectors of society pay more and more attention to the technology in the field of public security.At present,in order to maintain the safety of public environment,video monitoring equipment installed in streets,schools,communities,subways and other public places is increasing year by year.Gradually,the traditional video monitoring system is unable to process the growing large amount of video data effectively.Therefore,the development of intelligent monitoring system that can automatically detect,identify and alarm is of broad and profound significance to the maintenance of public security and the development of artificial intelligence.Abnormal event detection is an important part of intelligent video monitoring system,which mainly detects and warns abnormal events in monitoring scenarios to reduce the damage of abnormal events.In this paper,the problem of abnormal event detection in video monitoring is discussed and studied.First,the basic theory of abnormal event detection is introduced.Then,the research progress and methods of abnormal event detection are analyzed and summarized,and anomaly event detection models based on deep learning algorithm are proposed.The research contents and results of this paper are listed as follows:Firstly,when people flee in panic,they move around faster to avoid danger.Based orn the feature,a sparse auto-encoder network based on SSIM was proposed to detect crowd abnormal events.In this method,SSIM features are extracted to represent the change of the scene,and more meaningful information in SSIM features is mined by the sparse auto-encoder network.Finally,abnormal events are detected by Mahalanobis distance.Experimental results show that this method can achieve better detection results.Secondly,compared with normal events,abnormal events have great differences in appearance characteristics and motion patterns.Aiming at those features,a method based on the cascading deep net of appearance model and motion model is proposed.Gray images and optical flow graph of monitoring video data were extracted as the appearance representation and motion representation of data respectively.By using residual auto-encoder network,the appearance model and motion model is established to capture the characteristics of the normal events.In the abnormal detection,the data that matches the appearance model are firstly detected as normal,and the data that does not match the appearance model are initially determined as suspicious data.Then,the suspicious data matching the motion model are detected as normal,and the suspicious data not matching with motion model are finally judged as abnormal.Experimental results in multiple data sets show that the proposed algorithm can achieve better detection results.