Research Of Static Crowd Scene Counting Via Deep Network With Multi-branch Dilated Convolutions

Crowd density is one of important considerations in public security management.Excessive crowd density may be a major security risk because it is difficult to control.Crowd density estimation refers to estimating the crowd distribution in static scenes,which includes not only the simple estimation of crowd counting,but also the crowd density distribution map.It has become a hot research topic in the field of computer vision.The existing algorithms of crowd density estimation mainly use the target detector based on the deep convolutional neural network.Most of them focus on crowd counting,ignoring the quality of density distribution map and failing to make rational use of spatial distribution information of the crowd in the image.In this paper,a static scene crowd density estimation network was designed based on multi-branch dilated convolution block.By leveraging this network,a joint task of estimating the crowd count from a static single image and generating high-quality density map was realized.The main work of this paper for crowd density estimation in static scenes is as follows.First,an effective static scene crowd density estimation network is proposed,which integrates the advantages of detection and regression strategies and thus has good adaptability to crowd changes.At the same time,multi-branch extended convolution blocks are used to aggregate contextual information of different ranges,which makes it perform well in both the environment with large target size and sparse crowd density,and the highly-crowded crowd flow.Second,the two most mainstream indexes used in density map evaluation are peak signal-to-noise ratio?PSNR?and structural similarity index?SSIM?,having the problem of ignoring spatial related information.To address this problem,we introduce a new evaluation metric--spatial adjustment mutual information.It contains higher level spatial information,making the density map evaluation more objective.Finally,we carried out experiments on two public datasets?ShanghaiTech dataset and the UFC_CC₅0 dataset?.The results show that the proposed method outperforms the previous methods in almost all evaluation indexes,and has a higher accuracy.Meanwhile,the newly introduced evaluation index is verified.The experiment shows that it can better reflect the spatial distribution information of the crowd in the image,and the result is more objective and practical.