Research On Abnormal Event Detection Method In Surveillance Video Based On Frame Predictio

Video surveillance is applied in all kinds of public places,which improves the efficiency of disaster prevention and emergency response,and case detection,and guarantees the safety of people’s lives and property.However,traditional video anomaly detection requires long-term and rigorous observation of massive monitoring data to find abnormal events.This detection method is tedious and easy to cause fatigue,leading to the omission of key information.In order to find abnormal events accurately and respond to them timely,it is becoming more and more important to design an efficient and intelligent video anomaly detection method.Because of the occurrence and diversity of abnormal events,it is difficult to collect abnormal event data.In addition,due to the ambiguity of the definition of abnormal events,the definition of the same behavior depends on the specific scenario.Therefore,video anomaly detection is still a challenging task.This paper studies the detection of abnormal events in video,and its main contributions are as follows:(1)To solve the problem that the existing anomaly detection methods are unable to extract the time information between frames effectively,a prediction network Ind RNN-VAE that fuses an independently recurrent neural network(Ind RNN)and variational autoencoder(VAE)network is proposed.In the proposed prediction network Ind RNN-VAE,the spatial information of video frames is firstly extracted with VAE,then the latent features are obtained by a linear transformation.Secondly,the latent features are used as the input of the Ind RNN to get the temporal information between video frames.Finally,the latent features and temporal information are concatenated by residual block and input to the decoding network to generate the prediction frame.By testing on UCSD Ped1,Ped2 and Avenue public datasets,experimental results show that compared with the existing anomaly detection methods,the method based on Ind RNN-VAE has the performance significantly improved,and has the AUC values reached 84.3%,96.2%,and 86.6%,the EER values reached22.7%,8.8%,and 19.0% respectively.(2)An anomaly detection method based on a weighted prediction network WTAN is proposed,which solves the problem that existing anomaly detection methods cannot focus on the moving region where abnormal events occur and are easily interfered with by background information.This paper combines the triplet attention mechanism with the convolutional autoencoder to focus on moving objects in video frames,so as to improve the quality of video frame prediction.In addition,the three-frame difference method is used to extract the contour information of the moving target,and the weight matrix is obtained.On this basis,the weighted loss function is designed to give a higher weight value to the moving target,so as to improve the accuracy of anomaly detection.Experimental results on UCSD Ped1,Ped2,and Avenue datasets showed that the AUC and EER values achieved83.7%/23.4%,96.1%/10.2%,and 86.6%/20.9%,respectively.Compared with the existing anomaly detection methods,the method based on WTAN has the performance significantly improved.