Neural Network Method For Image Dense Crowd Counting Considering Plot Types And Perspective Rules

Crowd counting problems range from early scholars using computer vision to extract features,distinguishing and counting through expert knowledge,and using machine learning to return the number of people.Until now,with the development of deep learning and neural networks,neural networks have been used as crowd feature extractor of the image returns the number of people.Although the counting of dense crowds has made great progress in counting accuracy,it has ignored the fact that crowds still exist in geographic space and are constrained by geographic knowledge.This study attempts to place the dense crowd counting problem in geographic space and use geographic knowledge to constrain the dense crowd counting model in order to achieve the effect of further improving the counting accuracy.Based on this idea,the main content and conclusions of this article are as follows:(1)Construction of dense crowd count database containing perspective rules.This paper analyzes the existing dense crowd database and finds that it lacks perspective characterization in the dense crowd image scene.Such problems cause the dense crowd counting neural network to fail to learn the correct dense crowd characteristics under the influence of perspective rules,which in turn affects the counting accuracy.At the same time,it is found that the lack of a unified definition of population density in the open source database has led to the existence of crowd images with a large number of people in the database.There are also huge differences in the characteristics of dense crowds in the images with huge differences in numbers,which caused the shock phenomenon of the dense crowd counting neural network during the learning process.Eventually affect the counting accuracy.This article uses a combination of camera space and geographic space to describe perspective rules.This article refers to the definition of crowd density and gives a definition of crowd density suitable for the construction of a dense crowd database.Adding crowd range calibration to the database allows the crowd counting neural network to better learn crowd feature rules.(2)A dense crowd counting model based on residual neural network.In this paper,a new dense crowd counting network VGG-Res Next is constructed by fusing VGG network and residual neural network.The experiment proves that the network has greatly improved the counting accuracy.(3)Integrate the rules of geographic knowledge with the implicit rules of neural network.In this paper,the problem of counting dense crowds is placed in geographic space and constrained by geographic knowledge.The range of the crowd in the crowd image is restricted by the type of plot within the coverage of the camera,thereby reducing non-crowd noise.This method reduces the amount of model calculation and makes the model achieve faster convergence speed.In this paper,the image space and the geographic space are mapped to more accurately depict the perspective rules in the crowd image,so that the dense crowd counting neural network can obtain more realistic crowd feature rules.Experiments show that combining geographical knowledge can greatly improve the counting accuracy of dense crowds.