Video Anomaly Detection Based On Future Frame Prediction

In recent years,with the continuous development of society and econ-omy,the construction of smart cities has become the goal of all countries in the world,and smart security,as an important part of smart cities,is also the focus of current countries' construction.Therefore,video anomaly de-tection,as a core component of smart security,has aroused great attention from academia and industry at home and abroad.Video anomaly detection determines whether abnormal events occur by automatically analyzing be-havior information in the video.The research of video anomaly detection algorithms can greatly liberate manpower,and has very high research sig-nificance and application value.In the actual surveillance scene,there are still many problems with the existing video anomaly detection algorithms.The main manifestations are as follows:(1)In the surveillance video,because the distance from the camera is different,the objects in the picture show completely different scales.It is difficult to learn objects at different scales at the same time.(2)In the surveillance video,most of the areas are backgrounds that have noth-ing to do with abnormal events,and the area occupied by moving objects is only a small part.How to deal with such sparse scenes is the difficulty faced by current algorithms.(3)There is a lot of background noise in the video frame.The background noise is not helpful for judging the abnor-mality,but it will greatly interfere with the model's correct learning of the information of the moving object,which will greatly limit the performance of the related algorithms.The research content of this subject is how to solve the above problems.This paper proposes a video anomaly detection algorithm based on video frame prediction,which effectively improves the performance of the algorithm by eliminating or alleviating the above prob-lems.The main research contents and innovations of this article include two parts:1.A video anomaly detection algorithm based on residual future frame prediction network is proposed.By analyzing the characteristics that the background information of the surveillance video remains unchanged,a residual connection is added between the encoder and decoder of the de-signed network.The residual connection allows the background infor-mation to be passed into the decoder in advance,so that the predictor can be more focus on prediction of moving objects.And the residual connec-tion can make the gradient easier to pass into the encoder,which can make the encoder more fully trained.In the predictor,a GRU module based on dilated convolution is designed to predict the behavior of different scales through the dilated convolution of different receptive fields,which can pre-dict the behavior of different scales more accurately.Extensive experi-ments have been conducted on the CUHK Avenue,UCSD Pedestrian and ShanghaiTech datasets,and the results fully prove the effectiveness of the proposed method.2.A video anomaly detection algorithm based on multi-scale frame prediction is proposed.By designing a multi-scale frame prediction net-work to simultaneously predict objects of different scales in a video frame,the design of the multi-scale prediction network can also allow background information to be passed into the decoder in advance,which is very bene-ficial for sparse data learning.At the same time,in order to cope with the interference of noise on the frame prediction model,a perceptual loss func-tion is introduced to effectively mitigate the impact of noise by calculating the loss in a high-dimensional space with high semantic information.Ex-tensive experiments also have been conducted on the CUHK Avenue,UCSD Pedestrian and ShanghaiTech datasets,and the results fully prove the superiority and effectiveness of the proposed method.Notably,the pro-posed method obtains the frame-level AUC score of 87.8%on the CUHK Avenue dataset,which outperforms existing state-of-the-art approaches and achieves a new state-of-the-art.