The Analysis Of Crowd Density In Public Places

In recent years,with the improvement of living standards and transportation,the exponential growth of the world population and the consequent urbanization have led to more frequent gatherings of people.In this case,the crowd counting and density estimation is critical to building a higher level of cognitive ability in crowded scenarios such as crowd monitoring and scenario understanding,and is significant for the public safety arena.The analysis of crowd density is intended to count the number of people in a crowded scene and estimate density.Where the density estimation is intended to map the input crowd image to its corresponding density map.But like any other computer vision problem,crowd density analysis faces many challenges,such as occlusion,background interference,perspective distortion,scale variation,high levels of confusion,uneven distribution of people,uneven illumination,changes within and between scenes.The main work of this paper includes three aspects:Firstly,a network structure FDCNet for population density map and crowd counting for high-density diverse scenarios is proposed.The network integrates the feature maps of the lower and upper layers of the backbone,adapting to the scale changes caused by the perspective effect and sharing more features.It can eliminate background interference.The SE module is introduced as the channel attention module,considering the weight of the channel to selectively enhance the fusion of characteristics of the human head regions of different scales;A set of dilated convolutions is used as the end of the network,which increases the receptive field while ensuring less parameter quantities,including more detailed spatial information and global information to generate high quality density maps.Secondly,an improved network structure SWSNet is proposed to solve the FDCNet's ignoring the continuity of the image scale change.A scale weighted feature extractor is proposed to learn how to weight each single pixel to fuse multi-scale feature maps,so that each position in the image uses the correct context information;The subnetwork of the foreground segmentation of the crowd head is added.Learning more largescale salient features reduces the neglect of the shape characteristics of sparse populations and eliminates background interference.The relevant experimental results show the effectiveness of the improved structure.Finally,related experiments in four common datasets have fully verified the accuracy and robustness of the proposed network structure and the effectiveness of the improved structure.Furthermore,based on the population density map and population number generated by network prediction,a variety of simple but practical application methods are proposed.This includes sub-regional population estimates,density grading,population trend analysis,and high-density population testing.These analyses can be applied to high-density detection and early warning of crowd-intensive scenes in the field of public safety,effectively preventing the occurrence of security accidents such as group chaos and even trampling,and have high practical application value and prospect.