Abnormal Event Detection In Surveillance Video Based On Multi-Target Tracking

With the development of information technology and the decline of the price of hardware facilities,video surveillance system is widely used in banks,shopping malls,traffic intersections and so on,it has a significant effect in maintaining the stability of social security.Most of the early video monitoring system belongs to manual monitoring,because of various defects such as false alarm,slow processing efficiency and high human resource cost,they have been unable to meet the social demand.Therefore,video intelligent analysis technology came into being and made automatic monitoring possible.A system that uses this technology is called an intelligent video surveillance system,and it consists of many functional modules,such as target detection,target tracking and abnormal event detection,among which abnormal event detection is a very important one.Strengthening the research on abnormal event detection technology and continuously improving its work efficiency and detection accuracy is of great significance to maintain social and public security.At present,the development of abnormal event detection technology is relatively lagging behind,and there are still some problems in practice,such as the multi-scale of target object,the complexity of video scene and the diversity of anomaly events.This thesis carries out research work on the above problems,the main work of this thesis includes:1)Aiming at the scale change of the target object caused by different shooting distances and different shooting angles of the surveillance camera,a multi-scale target detection network model is proposed.Supplemented by feature pyramid structure and attention mechanism,it can make each feature map with different scales have strong semantic information and location information,and improve the target detection accuracy on the premise of increasing less computation;2)Aiming at the complexity of video scene and the diversity of events,an anomaly discrimination method based on target tracking and trajectory analysis is proposed,which can locate and track the position of target in video,and then identify and subdivide abnormal events through trajectory discrimination algorithm,which has good scene adaptability.The algorithm proposed in this thesis has been implemented and applied in the system platform of the intelligent video monitoring project of the middle route of the South-to-North Water Transfer Project.The algorithm uses the technologies of target detection and target tracking in deep learning to analyze and process the monitoring video in real time.Instead of relying on manual work,it can automatically find abnormal events from the video image,reducing the human resource cost.It improves the detection performance and enhances the credibility of the intelligent video monitoring system.