Research On The Method Of Counting People Based On Neural Networks

With the advancement of urbanization,China’s population is accelerating to gather in a few areas.And currently,the epidemic has been effectively controlled,and the number of people traveling will also increase significantly,leading to large-scale personnel intensive activities.While bringing about economic development,it also increases the difficulty of social security,urban traffic scheduling,and other aspects.It is an urgent need to analyze and count crowd activities in real-life scenarios such as tourist attractions,campuses,workplaces,and transportation hubs.Object detection and recognition is an important topic in the field of computer vision technology.Therefore,the use of crowd counting algorithms for security research has been a hot topic in recent years.However,due to issues such as background clutter and varying pedestrian scales in crowd images,the performance of current crowd counting algorithms still faces significant challenges.In recent years,research on the number of people counting has achieved certain results in terms of accuracy,running speed,and robustness compared to traditional number of people counting.However,most of them cannot balance the two.They have good applications in certain specific environments,but cannot meet all the requirements,In response to these issues,this article conducted a series of research population statistics based on the YOLOv5 algorithm.The main work is as follows:(1)In order to improve accuracy,this article proposes the introduction of attention mechanism,including ACE module and Shuffle Attention module.The experimental results show that the improved network model has higher recall rate and accuracy.(2)In order to accelerate the running speed,this article proposes the introduction of lightweight modules,including Shuffle Net V2 module and Mobile Net V3 module.The experimental results show that the improved network model can significantly improve the running speed with almost no change in accuracy.(3)By comparing the attention mechanism and lightweight module through experiments,it was found that the ACE module and Mobile Net V3 module performed better,respectively.Therefore,a combination of the two was proposed.Therefore,the Mobile Net V3 module was introduced into the YOLOv5 backbone network and the ACE module was introduced into the neck network.Through experimental comparison,it was found that using the improved network,compared with the original network,increased running velocity by 20 frame per second in the same GPU environment,At the same time,the accuracy of population counting was improved by about 5 percentage points.