Anomaly Detection With Weak Supervision In Surveillance Video Scene

Video anomaly detection technology can find suspicious or abnormal situations from a large amount of video data,and report warning timely.In this paper,under the condition of weak supervision,namely,given the test video that contain anomalies.When anomalies are found,the time the anomalous event occurs will be located.In view of the problems of the detection of fire,explosion,and traffic accidents in the surveillance video with weak supervision and the difficulty in precise location of the occurrence of abnormal events,this article does the following work:(1)In view of the problem of low video anomaly detection rate in surveillance video with weak supervision,an anomaly detection and coarse positioning method based on MSC8-3D is proposed.MSC8-3D is a frame-level supervised deep three-dimensional anomaly detection network,and the training samples with weak supervision only have video-level labels,so a frame-level pseudo-label generator proposed is trained to convert the video-level labels of abnormal videos into frames level pseudo-labels,and then train MSC8-3D network based on frame-level abnormal pseudo-labels and normal video frame-level labels.The results show that,on the UCF-Crime dataset,compared with the anomaly detection method based on deep multi-instance sorting,the AUC of anomaly event detection is increased by 0.2%,which proves the feasibility and effectiveness of the anomaly detection method.Secondly,based on anomaly detection on the video,coarse positioning is performed on the anomalous video.According to the characteristics of occasional and sudden abnormal events,some abnormal targets will eventually stabilize in the background,and the dynamic background subtraction is used to separate the front background from the video,and the separated background video is passed through the trained MSC8-3D The network extracts the features of the video clips,and finally uses unsupervised clustering to obtain the rough location of the abnormal occurrence time in the abnormal video.This method achieves the rough location of the abnormal event without using the frame-level label.The results show the feasibility of proposed method in the coarse positioning of abnormal events,the recall rate of coarse positioning of abnormal events increased by 0.6%,which proves the effectiveness of the method.(2)Aiming at the problem of the layer of MSC8-3D network,is too shallow,in order to get better video representation,anomaly detection and coarse positioning method based on MSRes35-3D is proposed.Among them,the MSRes35-3D network deepens the number of MSC8-3D network layers to 35 layers,which enhances the expression ability of the network.By introducing short residual connections,the shallow target features are preserved.Experimental results show that compared with the MSC8-3D-based abnormal event coarse localization method,the recall rate of abnormal event coarse localization on the UCF-Crime dataset is improved by 1.2%,indicating that the deep network learned Advanced semantic features have a better effect on coarse positioning of abnormal events.(3)On the basis of coarse positioning of abnormal events,a precise positioning method of foreground video based on coarse positioning strategy and abnormal target tracking is proposed.First,use the video background features learned in the method in Chapter 3 to perform clustering to obtain abnormal video frames after coarse positioning.Secondly,anomaly target detection is performed on the background of the abnormal video frame after coarse positioning to obtain candidate anomaly target positions.Then,by using the foreground information of the video,the characteristics of the initial position where the abnormal frame occurs can be captured accurately,and the background information can be exchanged.Finally,the frame backtracking method is used for the video foreground frame,which transforms the video anomaly detection task into determining the time point when the tracking target disappears,and realizes the accurate positioning of the start time and end time of the video anomaly event.The effectiveness of this method is verified through experiments,and the result of locating anomalies is improved.