Anomaly Detection Algorithm Of Intelligent Surveillance Video Based On Autoencoder

With the development of social informatization,a large number of surveillance videos appear.How to detect the anomaly from the mass surveillance videos in real time and conduct early warning is a hot topic in current research.At present,it is the mainstream of research to adopt the video anomaly detection method of unsupervised learning method,use deep neural network to carry out end-to-end feature learning on normal video data,and mark videos that do not conform to normal features as abnormal.Variational Recursive Neural Network(VRNN)is an unsupervised anomaly detection model.It combines Variational Autoencoder(VAE)with Long Short-Term Memory(LSTM).Each time step contains a VAE,and the prior information of VAE is provided by the preceding video frame through LSTM.Therefore,the model can directly model continuous video frame,and then predict future frames of input data,and finally distinguish the anomaly by using the prediction error.However,KL divergence tends to disappear during model training,resulting in the inability to learn meaningful underlying variables.In order to solve the above problems,two video anomaly detection algorithms based on frame prediction are proposed by adjusting the weight distribution between the reconstruction term and the KL term in the objective function.The main research contents are as follows:(1)In order to solve the problem of low accuracy of video anomaly detection due to the disappearance of KL divergence during VRNN training,KL cost annealing algorithm is introduced,and a VRNN based on KL annealing is proposed,called Annealing Variational recursive neural network(Annealing-VRNN).By adding a variable weight (4))from 0 to 1 to the KL term,it gradually anneals to a normal VRNN as the number of training increases.Experiments were carried out on public data sets ped1 and ped2.The results show that compared with common VRNN,the AUC values of the improved Anneal-VRNN are 94% and 96.56%,respectively,and the accuracy of video anomaly prediction is higher.(2)Although Annealing-VRNN alleviates the disappearance of KL divergence,it still falls into the disappearance of KL divergence with the increase of training times.Considering the influence of the actual KL value,the PID Control algorithm is introduced and improved,and a VRNN based on the improved PID control algorithm,Control-VRNN,is proposed.The model controls the KL value within the specified range,which completely solves the disappearance of KL divergence during the training process of the model.The results show that the AUC value of the improved Control-VRNN on ped1 and ped2 is significantly higher than that of VRNN,96.4%、96.42%,respectively.(3)The PSNR was used to replace the weighted sum of L1,MSSSIM and GDL loss to evaluate the difference between the predicted frame and the real frame,so as to identify anomalies.The experimental results show that the PSNR can forecast better assess the quality of the frame,improve the effect of anomaly detection model.In this paper,aiming at the problems of KL divergence disappearance generated by VRNN models during training,we have introduced KL cost annealing and PID control algorithms,and put forward two algorithms for video anomaly prediction.The effectiveness of the algorithms is verified by experiments,which has certain theoretical significance and application value.