| The continuous improvement of people’s income level and quality of life with the development of social economy,the crowd crowding scenes caused by people going out are also increasing,especially in urban business centers,transportation hubs,sports venues and tourist attractions during important events.In today’s increasingly complex social environment,the task of public security is particularly important.Among them,the research and analysis of public crowd aggregation behavior has become an important research topic in the field of intelligent image and video analysis.This paper mainly studies three parts: crowd counting,pedestrian detection,public gathering detection and decision algorithm.The main contents are as follows:In crowd counting,a multi-scale feature fusion population density estimation algorithm based on residual network is designed with the method of crowd density graph regression.At the front end of the network,the improved Res Net34 network was used to extract the crowd features,and at the back end,the three-array cavity convolution module with different void rates was used to capture the multi-scale features of the crowd,which increased the receptive field without changing the size of the feature map,enabling the network to obtain more scale details and edge information.The network has been trained and tested on Shanghai Tech and UCF＿CC＿50,respectively.The results are better than the current mainstream population density estimation methods,and the MAE(Mean Absolute Error)value obtained by the test is 61.2% lower than that of MCNN.Compared with SANet network,the performance is reduced by 6.4%,which verifies the feasibility and effectiveness of the network.In pedestrian detection,a lightweight attention mechanism pedestrian detection algorithm based on ACSP(Adapted Center and Scale Prediction)is designed.The feature extraction network of the algorithm adopts Res Net101 deep residual network,and the lightweight Channel Attention module ECA(Efficient Channel Attention)is introduced in each residual module.In the post-processing part of pedestrian detection,dou-NMS replaces the original NMS(None Maximum Suppression)to deal with the occlusion problem.The algorithm was trained and tested on the Cityperson pedestrian test set,and the performance of the algorithm was significantly improved compared with the original ACSP.The test results on the reasonable data set and partially blocked data set were 0.3%different from that of the APD algorithm using the DLA-34 backbone network.However,the test results on the severely heavy and bare data sets,All of them reach the optimal of the current mainstream pedestrian detection algorithm,which verifies the effectiveness of the algorithm.Decisions in the public gathering testing ways,in order to be able to detect real-time gathering for the public,as the starting point at the center of the crowd,a public gathering behavior detection algorithm is proposed.The algorithm uses YOLOv5 universal target detection network as the detector,and combines pattern recognition density peak clustering algorithm to effectively detect and determine pedestrian and crowd centers.First use target detection network to pedestrian detection of input image or video,to the pedestrian center as a data point calculation of pixel distance between each pedestrian to peak density clustering,clustering center by density compared with the preset threshold,determine whether the gathered crowds too dense,implementation of image detection,and the aggregation behavior of judgement.Then,the pedestrian detector designed in this paper is used to replace YOLOv5,which improves the performance of the algorithm in the case of serious pedestrian occlusion.Finally,experiments on three public datasets PETS2009,UMN and Malldataset verify the validity and reliability of the proposed method.Accurate crowd counting,high-precision pedestrian detection and reasonable public gathering detection play a positive role in social security prevention and control. |
