Research On Crowd Density Estimation Method Based On Convolutional Sparse Support Estimation Network

In recent years,with the rapid development of China’s economy,people’s travel and social activities have been increasing,and large-scale sports events and performances have also been increasing.Therefore,in some public places,such as stations,shopping malls,cinemas,etc.,a large number of people often generate,leading to problems such as stampede accidents,crowd conflicts,and traffic congestion.On the basis of deep learning,using convolutional neural networks to quickly estimate population density can effectively improve the accuracy of population density estimation.However,under the influence of factors such as large-scale and complex backgrounds,existing algorithms still face many difficulties and challenges.Therefore,this thesis mainly focuses on how to further improve the accuracy of population density estimation.In order to improve the performance of crowd density estimation in surveillance videos,this thesis proposes a crowd density estimation method based on sparse support basis estimation networks.First,in response to the problem of the areal feature of the crowd are not clear due to the interference of background noise,this thesis proposes a target awareness attention mechanism based on the sparse support basis estimation network of target characteristics,which can further enhance the Areal feature of the crowd and distinguish the crowd area from the background area.Secondly,in response to the problem of varying head scales caused by different distances and densities of people in the scene,this thesis proposes a target perception attention mechanism based on a sparse support basis estimation network of target scales,which can generate adaptive head scale perception rates and control the convolutional sampling mechanism to improve the distinguishability of heads at different scales.In order to verify the effectiveness of the method proposed in this thesis,experimental validation was conducted on standard datasets ShanghaiTech＿A,ShanghaiTech＿B,UCF＿CC50,and UCF＿QNRF.The MAE and MSE on the datasets ShanghaiTech＿A,ShanghaiTech＿B,UCF＿CC50,and UCF＿QNRF reached 51.2%and 86.8%,5.3%and 8.9%,194.4%and 276.2%,84.2%and 144.2%,respectively.On the ShanghaiTech＿A dataset,our method reduced MAE and MSE by 4.9%and 8.8%,respectively,compared to the most advanced PSGANet method.On the ShanghaiTech＿B dataset,our method decreased MAE and MSE by 1.3%and 0.8%,respectively,compared to the most advanced PSGANet method.The experimental results fully demonstrate the effectiveness of our method.