Research And Implementation Of Crowd Counting System Based On Deep Learning

In recent years,with the diversification of people's activities,the phenomenon of crowd congestion gradually appears in public places such as waiting halls and shopping malls.The higher the flow density,the more difficult the public management will be.At the same time,it will also bring safety problems such as stampedes.Crowds bring not only public administration and public safety problems,but also public health problems.As epidemic prevention and control becomes more and more normal,social and public health issues are getting more and more attention.In order to effectively prevent the spread of the epidemic or the occurrence of clustering,people should be prevented from gathering.Based on the above problems,thesis proposes a method combining the technology in the field of deep learning with the population statistics in order to realize a set of appropriate population statistics scheme by using the relevant computer vision model.The main work of thesis includes:Firstly,compare common people counting methods at present,based on the test method is applicable to this article real scenario requires accurate statistics crowd scenes,and regression method of density diagram is for the man's head is more sensitive,in view of the population density is very big scene difficult to detection the way people,can be used to return to figure way to estimate the population density.Therefore,the method based on detection is chosen in thesis.Secondly,Three algorithms with strong real-time performance in object detection field,YOLO,SSD and Retina Net,were selected for comparison,and the average detection accuracy of the three algorithms and the single sheet detection time were weighed.The YOLOV3 series algorithm was selected from the currently commonly used detection models as the recognition model of the crowd counting system in thesis.Thirdly,Introducing attention mechanism based on YOLOV3.Combining the research results of attention mechanism in the direction of computer vision and the application scenarios in thesis,channel attention module and mixed attention module are fused.In addition,the VOC-Person data set was extracted from VOC data set,and the content of Mall data set was fused to form VOC-Person data set,which was used as the data set of the training model in thesis.Fourthly,Based on the algorithm selected in the experiment in thesis,the counting module is integrated to build a crowd counting system using Python language,and the optimal model constructed is deployed into it.The crowd counting module is the core module of the crowd counting system.The crowd detection is carried out by deploying the optimal model and then the number of people is counted by the counting module.