Research On Crowd Counting And Crowd Abnormal Behavior Detection Methods

In recent years,crowd security issues in public places have also become a global concern and an urgent problem to be solved.The traditional video surveillance processing technology relies mainly on human resources for monitoring,which consumes a lot of financial resources and inefficiency,and can no longer meet the requirements of current intelligent video processing.At present,the method of analyzing video content based on computer vision technology has been developed rapidly.Generally,crowd analysis is carried out from two aspects: the number of crowd and the behavior of crowd.It has important academic value and application prospect of crowd counting and abnormal crowd behavior detection,and it is also one of the hotspots and difficulties in the field of computer vision.Counting the population in the crowd scene and estimating the crowd density,the crowd distribution information in the scene can be obtained to prevent safety accidents such as trampling due to overcrowding;to avoid sudden mass crowd incidents,abnormal crowd behaviors should be detected in video,and thus crowd danger signals can be found in time.Therefore,in order to prevent the occurrence of security accidents in public places where people gather,it is of great application prospect and academic value to conduct deep research on crowd density and crowd abnormal behavior detection based on video surveillance.Based on the deep analysis of the existing methods,this thesis constructs a deep generative adversarial network framework to estimate crowd density,and proposes an incremental non-negative matrix factorization method based on clusters constraints for the detection of abnormal crowd behavior.(1)In view of the problem of scale change in crowd scenes,a Deep Generative Adversarial Network framework is proposed for crowd density estimation and crowd counting.Firstly,the deep network is used to extract multi-scale abstract features of crowd scene images,and dilation convolution and deconvolution layers are used to normalize the feature resolution to maximize the reduction of redundant information while preserving the rich representation of the network multi-layer.Multi-scale representation of crowd scenes is obtained by fusing the multi-layer features of the network to improve robustness in term of crowd scale changes.The feature map is then further learned using a full convolution network to generate an estimated crowd density map.In order to make the error between estimated density map and the real gaussian density map smaller,this thesis uses the Euclidean loss and the perceptual loss respectively in the deep generative adversarial network to ensure that the estimated crowd density map is more realistic at the pixel level and the feature level,so that the estimated corwd density is more accurate.(2)Considering that there are fewer samples and unbalanced data distribution in the detection of abnormal crowd behavior,this thesis treats the abnormal behavior detection as a one-class problem,and Incremental Non-negative Matrix Factorization Based on Clustering Constraints is proposed for abnormal crowd behavior detection.The normal event is used as positive sample data,and the abnormal event is used as negative sample data.By finding an optimal positive sample cluster sphere in the feature space,it is judged as an abnormal event by mapping the test data to the position of the feature space.In the process of constructing the feature space,this thesis adds the two constraints of within-class and between-class to ensure that the positive samples within cluster is more compact inside,and the negative samples between cluster are more discrete,thus the feature space can be further optimized,and the accuracy of abnormal crowd behavior detection can be significantly improved ultimately.Finally,the crowd density estimation and abnormal crowd behavior detection methods proposed in this thesis to conduct crowd counting and anomaly detection experiments in various datasets,and the current popular research methods are compared in this thesis,which fully validates the effectiveness of the proposed method.