Research Of Abnormal Event Detection Based On Supervised And Weakly Supervised Learning

The traditional manual abnormal event detection is not only a great waste of human resources,but also unable to achieve the early warning or timely alarm of abnormal events which is not very reliable.Abnormal event detection based on deep learning means to automatically identify and locate dangerous events or abnormal behaviors in surveillance videos through deep learning algorithm,which is of great significance for ensuring public security.However,abnormal events detection is still a very challenging task due to the difficulties that both normal and abnormal events are not clearly defined,they are varied and cannot be exhaustively listed and the definition of normal and abnormal is different in different scenarios.In order to detect the abnormal events with large dangerous coefficient such as robbery,fighting and explosion in public places such as railway stations and squares,this paper proposes the following two algorithms to detect anomalies:(1)Supervised abnormal event detection algorithm based on YOLO v3+C3D.Compared with the unsupervised algorithm which only uses normal events for training,using both normal and abnormal events for training can better extract features which are helpful to distinguish normal and abnormal events.Therefore,a supervised algorithm based on YOLO v3+C3D is proposed.Firstly,the YOLO v3 target detection network is used to replace the traditional moving target detection algorithm to extract the foreground target more quickly and accurately,and to analyze the behavior from the point of view of the object.Combined with the pre-trained C3 D network,spatial and temporal features of the video data were extracted.Finally,normal and abnormal events were classified through the full connection layer and softmax function to achieve end-to-end anomaly detection.Good detection accuracy is obtained on UCSD datasets with relatively small scale.(2)Weakly supervised abnormal event detection algorithm based on I3D+LSTM.With the expansion of the dataset scale,the detailed labeling of training samples will become very complicated.Therefore,the multiple instance learning algorithm that only requires the video-level labels of normal and abnormal samples are used to reduce the dependence on label information.Firstly,the short-time sequence spatiotemporal features of training video were extracted by two-stream I3 D network,and the interframe motion information of the long-time sequence was further extracted by combining with LSTM.This network structure of short-time sequence plus long-time sequence modeling can make full use of time context information to obtain the characteristics of abnormal events.Finally,three layers of full connection layers are used to score each video clip,and the network is optimized by ranking loss.This algorithm achieves the highest anomaly detection accuracy on large scale datasets which named UCF-Crime and Shanghai Tech.