Crowd Abnormal Behavior Detection Based On Deep Learning And Sparse Representation

With the rapid development of computer hardware and digital.information technology,computer vision has become a hot research field and a widely concerned object.In recent years,considering the important role of public security incidents in national construction and development as well as public security,a large number of video surveillance devices have been installed in public places such as railway stations,subway stations,large shopping malls and streets.Traditional video surveillance technology has been unable to meet the real-time and effective monitoring of massive video.How to monitor and process video quickly,in real time,and intelligently becomes a hot research direction.Among them,abnormal behavior detection is an important research content in the field of intelligent surveillance.It mainly completes the detection of abnormal behavior in video surveillance,as well as the prediction and early warning,so as to reduce the harm of abnormal events to public security.This paper focuses on the study of abnormal crowd behavior detection.The main work and achievems of this paper are as follows:Firstly,this paper completes the extraction of video motion features,introduces in detail the principles of the three algorithms for moving target detection,and discusses the advantages and disadvantages of the three algorithms.Based on the idea that"abnormal events"are not"normal events",we propose a abnormal crowd behavior detection method by using sparse representation.The theoretical basis of sparse representation used in this paper is elaborated.Based on the traditional sparse representation algorithm,the dictionary update algorithm based on sparsity is adopted,which further improves the accuracy of abnormal behavior detection.Secondly,a sparse representation abnormal behavior detection algorithm based on spatial-temporal feature is proposed.Firstly,the histogram of optical flow(HOF)is extracted as the temporal features of the frame.Secondly,the saliency information(SI)of the image is calculated as the spatial feature.The spatiotemporal features of video information are obtained by fusing temporal and spatial features.The sparse representation anomaly detection algorithm based on sparse representation is compared with the traditional HOF feature sparse representation anomaly detection method.The sparse dictionary is trained by the features extracted from normal samples.In the detection stage,the sparse reconstruction cost is calculated to detect anomaly.Experiments show that this method can effectively improve the rate of video anomaly detection with low contrast.Thirdly,on the basis of anomaly detection with the fusion of spatial and temporal features,a sparse representation crowd anomaly detection method is proposed to improve the effectiveness of features and the accuracy of detection.On the basis of the original PCANet,pooling operation is adopted,which effectively reduces the dimension of the extracted high-level features,solves the difficulty of sparse representation classification,and avoids overfitting.The improved PCANet network is used to extract high-level features from the temporal and spatial features of video frames.to construct a sparse dictionary of normal behaviors based on high-level features.The high-level features of the video image to be tested are sparse reconstructed,and the reconstruction cost is used to judge whether the event is abnormal.In view of the difference of anomaly detection rate caused by the different number of improved PCANet network filters,a comparative experiment is carried out.The improved PCANet with appropriate number of filters is used to extract high-level features and compare low-level features.This paper conducts experiments in the UMN and WEB datasets,the experimental results show that the method can effectively detect abnormal behaviors in different scenarios.Compared with the traditional low-level features,the anomaly detection rate is improved effectively.