Study Of Video Anomaly Detection Based On Future Video Frame Prediction

With the development of computer technology,excellent technological products have emerged in our real life.Surveillance cameras are one of the most widely used products.The installation of security cameras in public can ensure the safety of public property and citizens as much as possible.Using surveillance cameras and the artificial intelligence technology,video anomaly detection is a hot area of research in computer vision recently.However,due to the low incidence of abnormal events and the small number of samples,the frequent changes of scenes and diversity of monitoring perspectives,it is still a great challenge to detect abnormal human behaviors in surveillance videos.In this paper,we focus on the field of video anomaly detection,and on the basis of summarizing the existing research work,we propose two new methods,namely an anomaly detection method via memory-augmented frame prediction and a dual-branch framework with convolutional attentive block for video anomaly detection.The main work of this paper are as follows:(1)An anomaly detection method via memory-augmented frame prediction is proposed,which aims to solve the task of anomaly detection by predicting future frame.Take the continuous video frame images as the model input,our model can automatically realize end-to-end learning.Specifically,the memory module defined in the generator network uses the learned normal patterns to produce high quality query features,thereby increasing the loss error between normal data and abnormal data.The introduced attention mechanism helps the network to readjust feature representations in terms of channels,which makes the model pay more attention to feature information in important channels.Experiments on benchmark datasets show that the anomaly detection in surveillance videos via memory-augmented frame prediction has superior performance.(2)A dual-branch framework with convolutional attentive block for video anomaly detection is proposed,which integrates the prediction branch and the reconstruction branch.The prediction branch utilizes video frames to learn the appearance features,and the reconstruction branch uses optical flows to learn the motion features,thus our model can learn information of abnormal events in both appearance and motion.The attention mechanism models the dependencies between feature representation channels.In addition,a convolution attention module is proposed to learn the global structure of local features of each branch,so that the context of the feature representation is enhanced.Moreover,the multi-scale memory modules are introduced to record the regularity of normal data at different scales,and guide the model to detect abnormal events better.Experiments on mainstream benchmark anomaly detection datasets show that the dual-branch framework with convolutional attentive block for video anomaly detection has better performance.(3)A new video anomaly detection dataset named Nut is proposed.Considering that the existing video anomaly detection datasets have a small number of sample perspectives and scenes,and abnormal events in real life are multi-scenes and multi-perspectives,a new large-scale video anomaly detection dataset named Nut is proposed in this paper.This dataset integrates a large number of samples from existing benchmark datasets and real surveillance videos,including 10 abnormal events that have significant negative effects on social life,including intrude,fight,across,discard,weapon,rush,accident,crowd,arson and tumble.The experimental results of popular anomaly detection methods validate the availability of Nut dataset.