Crowd Counting Based On Adaptive Map Refinement

The purpose of Crowd Counting is to predict the number of people in the monitoring scene and generate a density map to reflect the population distribution in the current scene.As the first and most important part of crowd management,automatic crowd counting can monitor the crowd density in the area and alert the administrator to perform security control when the density exceeds a specified threshold.Due to the exponential growth of the world’s population in recent years leading to urban centralization,collective activities have become more frequent.With such a large population gathering and flow,crowd counting and analysis have become particularly important.Like other computer vision tasks,crowd counting faces many challenges,including: severe occlusion of dense crowds,interference from complex backgrounds,uneven distribution of crowds,appearance changes inside and outside the scene,and scale variations and perspective issues and more.These challenges make this research topic extremely difficult,but it has also attracted a large number of researchers to invest in the in-depth research.With the rapid development of Convolutional Neural Networks(CNN),CNN-based counting models have sprung up and made significant progress.However,the common disadvantage of these models is that they tend to treat the entire image of different data sets equally with a single fixed structure,which cannot cope with various complex scenes with different population densities.Therefore,this paper proposes a crowd counting algorithm based on adaptive map refinement,which aims to use the unique representation capabilities of networks with different levels of complexity to process regions of different densities.The specific research work is as follows:(1)Through careful analysis of existing work and theory,the advantages and disadvantages of existing algorithms are summarized.By comparison,it is found that deep complex networks can handle high-density areas,while shallow simple networks can achieve better performance in sparse areas.Based on this theory,a corresponding solution for crowd counting is proposed.(2)A coarse network with multiple columns of shallow structure is designed,each column is composed of convolution kernels of different sizes,so each column has a different receptive field,which is used to enhance the ability of the density variations of different feature resolutions.Taking VGG16 as a prototype and combining Deformable Convolution,a fine network with a deep structure is designed.By adding the proposed adaptive fusion strategy,this network can selectively suppress noise and automatically focus on the size of the matching crowd,thus greatly improving the counting accuracy in complex scenes.The two networks are used to further verify the unique characterization capabilities of networks with different degrees of complexity.(3)This paper proposes a crowd counting algorithm based on adaptive map refinement to utilize the advantages of networks with different degrees of complexity.First,a novel attention mechanism called counting attention is designed,which can automatically locate the dense area of the input without prior knowledge.Then a fine network and a coarse network are used to separately process the dense and sparse areas in the scene to form the complete population density map required by the model.(4)A smooth network is designed to deal with the fragmentation of the generated density map caused by the splicing of two networks,and a novel GAN loss is proposed to make the generated density map more realistic.