Research On Group Counting Based On Deep Learnin

The rapid development of artificial intelligence will affect all aspects of human industrial production and social life profoundly.Developing artificial intelligence technology has been listed as a national strategy by the state council.Using artificial intelligence technology to build a smart city and create smart life has become an urgent need for the country’s modernization.Nowadays,the automation technology using computer vision has been widely applied to video surveillance,industrial automation and many other fields.As a hot spot in vision community,crowd counting has a very wide range of application scenarios in life,and accurate estimation of the number of objects in the scene is the key to this task.In stations with large human flow,the accurate estimation of the number of people in the scene can provide a basis for the scientific management of relevant departments.Automatic and accurate statistics of the number of relevant parts for industrial automation detection is very important.and real-time crowd density monitoring in large public places can help identify security risks and provide early warning.The goal of crowd counting is to utilize computer vision technology to estimate the number of group objects accurately in an image scene.Depending on the specific application scenarios,the methods and steps adopted by the crowd counting algorithm are also different.In this paper,the people counting based on deep learning and the locomotive running part bolt counting based on semantic segmentation are studied in depth respectively.According to the different characteristics of practical application scenarios,the problems of existing relevant algorithms are analyzed,and the counting algorithms for different application scenarios are proposed.The main research work of this paper can be summarized as follows:1.for the population counting problem in complex scenes,the traditional crowd counting algorithm based on detection and regression can perform well in the scene with relatively sparse crowd distribution,but it is difficult to adapt to the complex scene with severe crowd occlusion and large spatial scale variation.In recent years,the crowd counting algorithm based on deep learning has made great progress in counting performance,but most of the network structures are complex and the training steps are time-consuming,which brings difficulties for the practical application.In this regard,this paper proposes a new counting method MSF-CNN.Finally,experiments on several public data sets show that the proposed method can not only adapt to more complex counting scenarios,but also has the advantages of simple structure and convenience for training model.2.compared with crowd counting,bolt counting in industrial detection is a more complex problem,which not only requires accurate real-time statistics of bolt number from complex scenes,but also requires bolt classification,that is,counting with certain semantic information.To end,this paper proposes a novel based on semantic segmentation for locomotive direction of bolt count,to adopt semantic segmentation of the complex subway locomotive extracted from direction of different types of bolts,finally counts the bolts by using statistical techniques on the real collected dataset experiments show that the method proposed can not only better retaining bolt details and more accurate count,but also can provide a reference for locomotive bolt running stocks anomaly detection.3.On the basis of the aforementioned works,this paper designs and implements a crowd counting software system which includes two modules:crowd counting on the subway platform and bolt counting on the running part of the subway locomotive.This paper not only gives the functional structure of the software and explains the use of the software,but also shows the application effect of the software system.Overall,in this paper,according to different application scenarios colony counting method carried on the thorough exploration and research,combined with the characteristics of actual application scenarios and requirements,two new different types of counting method are proposed,namely the people counting model based on density estimation and model of bolt count based on semantic segmentation,and multiple public data sets is used to verify the validity of the model.Finally,based on the previous two works,a crowd counting software system based on deep learning is designed and implemented.