Research On The Detection Method Of Abnormal Crowd Behavior Based On Sequence Analysis

With the increasing demand of public safety prevention and the wide application of video surveillance technology,the detection of abnormal crowd behavior has become one of the hot research issues in the field of video analysis and understanding.Abnormal crowd behavior detection aims to model and analyze the behaviors of pedestrians in surveillance videos,distinguish the normal and abnormal behaviors in the crowd,and timely detect disasters and accidents.The traditional video surveillance technology can only transmit the surveillance picture on the display screen in real time,and cannot automatically analyze the video content.Therefore,the detection of abnormal crowd behavior in public places needs a lot of manpower,material and financial resources.In view of the traditional video surveillance methods in public places such as railway stations and airports are time-consuming,laborious and inefficient,it is of great significance to study an intelligent automatic analysis video surveillance technology,which can help staff to timely find hidden safety problems in public places and ensure the safety of staff in public places.With the rapid development of computer vision application technology,surveillance video content analysis technology has also made great progress.However,in the task of abnormal behavior detection,there are still some difficulties and challenges: 1)most of the existing methods pay attention to the spatial feature anomalies and ignore the temporal feature anomalies,which makes it easy for the algorithm to miss detection in the process of detection,resulting in poor algorithm performance;2)Due to the wide variety of monitoring scenes,complex crowd movement and changing crowd density,the detection algorithm is not ideal for the location of abnormal behavior area.To solve these problems,this paper takes deep learning as the carrier to carry out the research on the detection method of abnormal behavior in people based on sequence analysis.First of all,in view of the abnormal behavior detection studies generally use spatial features while ignoring temporal features to detect anomalies,resulting in abnormal behavior missed detection,a method of abnormal frame detection based on spatiotemporal prediction reconstruction is proposed.The Generative Adversarial Network is used as the framework,and the generator is composed of a prediction network and a reconstruction network in series to predict and reconstruct future frame.The discriminator is composed of Pix2 Pix GAN network,which is used to identify the authenticity of future frame.The prediction network is composed of U-Net network and LSTM network,which is used to extract the spatial-temporal features of video sequences and ensure the generation of pedestrian motion information in future frame.The reconstruction network is composed of UNet++ network,which is used to reconstruct the appearance characteristics of pedestrians in future frame.Finally,the Peak signal-to-noise Ratio is used to calculate the error between the future frame and the corresponding real frame to detect abnormal frame.Experimental results on public data sets show that this method can effectively detect abnormal behavior.Secondly,aiming at the problem of unsatisfactory location of abnormal behavior region,an abnormal behavior detection method based on self-attention enhancement is proposed.Based on the prediction network composed of U-Net and LSTM,the method integrates the Attention mechanism,which makes the prediction network pay more attention to the local significant features of the input images and suppress the features of the irrelevant regions,so as to improve the sensitivity and accuracy of the prediction network.In order to detect and locate abnormal behavior,two kinds of discriminators are trained simultaneously: the frame-level discriminator Pix2 Pix GAN and the pixel-level discriminator Patch GAN.Among them,Pix2 Pix GAN is used to judge the true and false of future frame;Patch GAN can be used to determine whether future frame are true or false,as well as to locate areas of abnormal behavior in real image.Experimental results on public data sets show that this method has good performance in abnormal behavior detection and location.