Research On Crowd Abnormal Behavior Detection Based On Deep Learning

In the field of security,cameras have an irreplaceable role,but the traditional security monitoring system needs to invest a lot of manpower to monitor and analyze the monitoring content and find abnormal situations.The crowd abnormal behavior detection algorithm studied in this paper aims to use computer technology to replace human to realize automatic detection and early warning of crowd abnormal behavior in surveillance video.Most of the traditional crowd abnormal behavior detection algorithms are based on manual features.Due to the nonlinearity and uncertainty of crowd abnormal behavior,these traditional methods have shortcomings such as poor robustness and weak generalization.The network model based on deep learning has a strong ability to fit nonlinear functions,so this paper proposes to use deep learning technology to solve the problem of abnormal crowd behavior detection.In this paper,we propose a probabilistic framework for detecting abnormal crowd behavior based on deep learning technology,which named Variational Abnormal Behavior Detector（VABD）.Our work mainly has the following three contributions:（1）In order to solve the problem that the features of video sequences are difficult to extract,and at the same time consider the influence of the context content of the video frame at a certain moment on the judgment of abnormal behavior,this paper designs a conditional variational autoencoder network structure for short video sequences.to extract features from the input data.（2）In order to make the video frames generated by the network and the input video frames have better motion consistency,this paper proposes a motion loss function based on an improved optical flow network,which fully considers the feature maps of different scales to generate results.The experimental results show that it can effectively improve the detection accuracy of the algorithm.（3）In order to further improve the performance of the detection model,the paper also applies the idea of Generative Adversarial Networks to the VABD algorithm framework,and at the same time,in order to eliminate the interference of the background and other factors in the video frame to the classifier,the paper also proposes a method based on The attention mechanism algorithm module of the feature layer.The experimental results of our proposed VABD algorithm on internationally pub lic datasets such as UCSD^[23],CUHK Avenue^[24],IITB-Corridor^[25]and Shanghai Tech^[26]show that it can effectively detect the video frames in the video sequence contain a bnormal behaviors.At the same time,the VABD algorithm proposed in this paper is c ompetitive with the most advanced algorithms in this field in terms of detection accur acy.