Crowd Counting In Complex Scenes Based On Multi-scale Fusion Approaches

With the improvement of the level of urbanization and the increasing size of the population,a large number of dense scenes hide many hidden safety hazards.Therefore,crowd counting research is very necessary and plays an important role in early warning planning,public management and other fields.In recent years,the rapid development of computer hardware technology and deep learning has led to outstanding results in the research on crowd counting algorithms based on convolutional neural networks.However,due to the existence of thorny problems such as the drastic change of the crowd scale and the cluttered distribution of the crowd in real scene,the existing algorithms still face severe challenges in counting performance.This dissertation conducts research on the algorithm structure and loss function,targeted solutions are proposed to solve the above problems and improve the performance of the algorithm in further.The main research work and innovations of this dissertation are as follows:1.Aiming at the problem of dramatic scale changes in actual scenes,a type of crowd counting method based on context-aware feature reaggregation is proposed.The method establish a feature enhancement module based on the atrous convolution to cap-ture multi-scale context information,and the fusion module is used to strengthen the se-mantic information in the multi-scale features;a multi-scale fusion structure is designed at the output to further enhance the method's scale perception ability.This method can not only improve the accuracy of counting,but also generate high-quality estimated density maps.2.Aiming at the problem of disorderly crowd distribution in complex scenes,a crowd density estimation method based on the different architectures of semantic ag-gregation and density aware is proposed.The cluttered distribution of the crowd in the real-world will directly affect the counting performance of the counting methods.En-hancing the spatial perception of the crowd density distribution can improve the quality of the density map generation,thereby improving the counting accuracy of the method.The proposed method establishes a feature fusion network and uses the attention mecha-nism to improve the expression of fusion features;In order to further improve the density perception ability of the method,a density aware module is established at the back of the structure,and the attention masks of different density levels are used to generate the corresponding regions' density map.This method can reduce the influence of the clut-tered crowd distribution on the quality of the density map,further improve the counting performance and generalization ability.3.Extensive experiments on five benchmark datasets for the two methods men-tioned in this dissertation.Firstly,extensive experiments on ShanghaiTech,UCF-QNRF and UCF_CC₅0 datasets show that the proposed methods have achieved better perfor-mance;Secondly,The ability and robustness of these methods are well verified on the two newly proposed large-scale JHU-CROWD++ and NWPU-Crowd datasets;At the end,additional ablation experiments further illustrate the effectiveness of these pro-posed modules.