Research On Video Anomaly Detection Based On Frame Prediction

With the development of software and hardware technology,video surveillance systems have become widely used.These systems are important for the maintenance of national security and social stability.Video anomaly detection is one of the core functions of the video surveillance system,however,it takes huge labor and time costs to find abnormal events completely by manual.Besides,the monitoring efficiency is low and it is easy to miss important information for monitoring staff.Therefore,detecting abnormal events in the videos automatically has become an important and urgent research topic.The research on video anomaly detection is conducive to the further realization of the intelligent video surveillance system,having important value for intelligent security.Aiming at the problems of rare abnormal samples and diversified types of abnormal events in practical applications,this paper only uses normal video data to train the model,and studies two video anomaly detection algorithms based on frame prediction.(1)Video anomaly detection method based on multi-space frame prediction.Aiming at the problem of single detection angle in the existing work,which is susceptible to misjudgment by detailed pixels,the method predicts future frames in the original image space and latent space,and detects abnormal events in multi-space.Specifically,based on a given historical video sequence,the method first obtains the predicted video frame and its encoding in the latent space via the predictive neural network and the coding neural network;then designs a variety of prediction error loss terms to train the network so that the model can learn the behavioral patterns of normal events in multiple spaces.In the detection stage,the method compares the predicted video frame with the real video frame,calculates the weighted score in the image space and the latent space for each frame of the video to be detected,and finally find the abnormal event according to the scores.This method can capture the distribution of normal events from different spaces,improve the quality of video frame prediction results,and effectively reduce the error rate of video anomaly detection.(2)Video anomaly detection method based on a spatio-temporal dual discrimination enhancement mechanism.Video sequences have spatial and temporal dimensions,and contain a wealth of motion information.Learning the spatio-temporal information of videos is conducive to identifying abnormal events.However,the existing works lack considering of the time dimension of the motion information,which limits the detection accuracy.To this end,the method first uses a generator of a generative adversarial network to predict future video frames and uses a neural network to estimate the optical flow between consecutive video frames.Then the spatial discriminator and the temporal discriminator are used to distinguish the authenticity of the input video frame and the input optical flow respectively.After adversarial training,the generator's ability to learn the spatial appearance information and motion information of the normal event is enhanced.Finally,the method realizes the detection of abnormal events by measuring the similarity between the predicted video frames and of real frames.This method can effectively capture accurate normal event behavior patterns from the two dimensions of time and space,enhance the model's ability to distinguish between normal events and abnormal events,and improve detection performance.In addition,because abnormal data often appear in local areas,this paper tries a patch detection strategy,which calculates the scores according to the local similarity between the predicted frames and the real frames,further improving the detection accuracy.To verify the effectiveness of the methods,experiments on three public data sets of different scales are conducted.Compared with existing methods,the methods in this paper have certain advantages under different evaluation metrics.