Research On Deep Learning Based Crowd Counting And Density Estimation

With the continuous growth of the population base,the active crowd not only has the risk of triggering public safety incidents,but also causes huge pressure on the city's public security and control.Therefore,the crowd counting analysis of complex public places can effectively reduce the risk of public safety events,and also provide some help for urban public security.However,with the in-depth research in the field of crowd counting in recent years,many problems have arisen,such as the diversity of scenes,the varying sizes of pedestrians,the mutual occlusion of the crowd and the different viewing angles,all of which can result in poor counting performance.As a result,it is difficult to implement research in this field and meet the needs of current real-world scenarios.Therefore,this article researches and analyzes the above problems,and proposes two new crowd counting schemes from the multi-scale perspective of the crowd:(1)A crowd counting model CFFNet based on feature fusion and encoding and decoding is designed.Firstly,the traditional classification network Inception V3 is improved to construct the front-end network module of the CFFNet model,the low-dimensional crowd feature information is retained through multi-output mode,and the crowd feature information of various dimensions and depths is obtained to complete the whole encoding process of crowd image.Then,on the basis,the back-end network module of the CFFNet model is jointly constructed by the proposed feature fusion module and feature decoding module,which completes the feature decoding and fusion of crowd feature information of different dimensions,and output the final estimated density map.Finally,the indispensability of the back-end network module is verified through comparative experiments,and the CFFNet model is tested on four benchmark crowd datasets to verify the effectiveness of the model and its adaptation to different crowd scales.At the same time,the expansion experiments are carried out by using crowd images outside the crowd dataset,which to a certain extent verified that the model has good generalization ability.(2)A multi-stage based lightweight crowd counting model MS-LWNet is designed.Firstly,the standard convolution and depthwise separable convolution are compared,and the problems still existing in the application of depthwise separable convolution to the field of crowd counting are analyzed.Then the multi-scale residual block is constructed through two parallel "bottleneck" modules and "short circuit" modules,and on the basis of this residual block,a MS-LWNet model based on "front-end body back-end" is constructed.Finally,the model was tested on the Shanghai Tech Part A dataset,and the size of the model parameters was only 5.19 MB,which provided the possibility for the deployment of the model on mobile devices,and the model also reached 83.6 in the MAE metrics,which further demonstrated the multi-scale residual block has a good ability to learn crowd characteristics,so as to achieve a balance between counting effectiveness and efficiency.In general,this paper takes the crowd dataset as the research base,the crowd counting and density estimation as the research goal,and solving the crowd scale problem as the research purpose.By constructing a crowd counting model based on crowd scale,and performing training and testing tasks on the crowd dataset,the problem of different crowd scales is effectively solved,and the effectiveness and generalization of the model are verified at the same time.