Research On Video Anomaly Detection In Surveillance Scene

With the gradual strengthening of the awareness of the public in the field of public safety and the rapid advancement of automation and intelligent technology,the science and technology for ensuring public safety have received more and more attention.At present,in order to more effectively maintain the security in public environments,many public places have installed video surveillance equipment.However,the traditional method of using manual observation for alarming has been unable to effectively handle the growing mass of video data,and it will cost a lot of manpower.Material resources.Therefore,an intelligent monitoring system that can automatically detect,identify,and alarm is of great significance to public safety.This article discusses and studies the problem of video abnormal event detection.First briefly introduce the basic theory of video anomaly detection,then analyze and summarize the research status and common methods of video anomaly detection,and finally put forward two anomaly detection models based on deep learning algorithms.The research content and results of this article are as follows:1.At present,some video abnormal event detection methods based on deep learning directly determine whether the test sample is a normal or abnormal event through the size of the reconstruction error of the autoencoder.However,in practical applications,the reconstruction error of some original abnormal test samples after self-encoding is also less than the set threshold,thus misjudged the test sample as a normal event,resulting in the situation of missed detection of abnormal events.In response to this shortcoming,this paper proposes an anomaly detection model that integrates an autoencoder and One-class SVM.This method first extracts the deep spatiotemporal features of the region of interest through a 3D convolutional neural network to represent motion events,and then uses the noise reduction autoencoder reconstruction error to detect the size relationship between the set threshold to exclude most abnormal events.class SVM once again eliminates abnormal events forevents whose reconstruction error is less than the threshold.2.Due to the complexity of the actual monitoring scene,different normal events are also likely to show a large difference.The data is characterized by multiple distributions,that is,the sample space has large variance.If a single hypersphere One-class SVM is used to describe the feature space may lead to the situation that some samples that are originally abnormal are detected as normal samples.To solve this problem,an anomaly detection method based on deep convolution clustering is proposed.This method first extracts fixed-size image blocks and optical flow blocks from the video and fuses them in a pixel-level manner to obtain a spatiotemporal information containing motion and appearance.Then use the deep clustering network to perform feature clustering while performing feature representation on the spatiotemporal information.Finally,by constructing a One-class SVM model for each clustering cluster,use it to detect abnormal events.