Anomaly Detection In Surveillance Videos Based On Deep Learning

Surveillance cameras are becoming more and more common in public places.However,the mere use of law enforcement agencies to detect anomalous events in surveillance videos will result in a large amount of waste of resources and inefficiency.Therefore,if the computer is able to perform online anomoly detection after receiving the video transmitted by the monitoring device,this will undoubtedly bring great help to the law enforcement department.But,to achieve online anomaly detection in surveillance videos,it is necessary to overcome the following three difficulties: meeting the real-time requirements;effectively utilizing the long untrimmed surveillance videos;addressing the complexity of the environment.Corresponding methods proposed to solve these three difficulties in this thesis:1)A mobile three-dimensional convolutional network is proposed,combined with the structure of the effective convolutional network,to design the mobile effective convolutional network(M-ECO)for action recognition,in order to achieve real-time algorithmic.The network can realize the recognition of long-term and short-term actions in the video.In processing 8 frames,the accuracy on the UCF101 can reach 92.1%,and the runtime reaches 31.3 VPS(Videos per Second).2)The above M-ECO replaces the feature extraction part in the UntrimmedNet combined with the attention weight mechanism,and the mobile-UntrimmedNet(MU-Net)is designed to effectively utilize the untrimmed video.A class activation sequence adjacent comparison is proposed.Combined with the improved network,the accuracy on the THUMOS 2014 can reach 81.9% without extracting optical flow information.In addition,for the temporal action detection problem,the mean average precision of the network on the THUMOS 2014 can reach 28.3 on the premise that the intersection over union is set to 0.3.3)Two additional data argumentation methods are used to enhance the robustness of the MU-Net to cope with the complex environment of cameras.In this way,the i MU-Net achieves a 46.3% accuracy on the UCF-Crime.In addition,an online video understanding algorithm is used to achieve online anomaly detection in video.