Research On Dense Crowd Counting Algorithm Based On Multiscale Perceptual Network

Crowd counting is the core technology of crowd monitoring and crowd analysis,which is of great significance for the scientific deployment of security measures in public safety management.In recent years,with the development of deep learning technology,a large number of deep neural networks based on density estimation have emerged for crowd counting work,but these networks are difficult to meet the needs of practical applications in terms of counting accuracy or detection speed.In order to improve the counting accuracy and detection speed of the crowd counting network,this paper designs and builds a dense crowd counting algorithm based on a multi-scale perception network.The multi-scale aggregation module is designed to solve the general problem of head scale in crowd counting,and atrous convolution is introduced to prevent the model from running slowly due to the large amount of network parameters;at the same time,a high-resolution generation network is designed as the network back-end to improve the resolution of the density map and further improve the counting accuracy of the model;Finally,the multi-scale feature extraction network designed based on the multi-scale aggregation module and the high-resolution density map generation network are connected and fused to build a dense crowd counting network based on the multi-scale perception network.The main research work of this paper is summarized as follows:Firstly,in view of the different head scales in crowd counting images and the large amount of parameters of crowd counting network,a multi-scale aggregation module introducing atrous convolution is proposed,and a multi-scale feature extraction network is designed based on this module.By sending the feature map to the multi-scale aggregation module for feature extraction multiple times,the extraction of human head features of different scales in the image is realized,and the method of combining features of different scales is added;The introduction of hole convolution to replace the traditional convolution achieves the purpose of covering the full scale.Secondly,a high-resolution density map generation network is proposed to solve the problem of low resolution and loss of details in the generated density map,which affects the accuracy of crowd counting.The purpose of generating high-quality density maps is achieved by repairing the details of feature maps with convolution kernels of different sizes,and transposing convolution to restore the resolution of feature maps.Finally,in order to take into account the speed and accuracy of crowd counting,combined with multi-scale feature extraction network and high-resolution generation network,a crowd counting algorithm based on multi-scale perception network is designed.Experiments based on the above-mentioned key technology research show that the multi-scale perceptual neural network adopts a multi-scale feature extraction network that introduces atrous convolution to extract features,which not only considers the size of the human head scale features,but also increases the combination of scale features.The amount of network parameters is greatly reduced.The model achieved an average absolute error of 67.7 and 8.9 in the ShanghaiTech dataset A and B,respectively,leading the mainstream crowd counting network,and the actual detection speed was 40%faster than the CSRNet network with a small number of parameters.