Research On Crowd Density Estimation Algorithm Based On Multi-scale Information And Attention Mechanism

In the past decade,with the growth of world population and the continuous improvement of living standards,large-scale population movements in holidays have become far more common,which consequently bring great challenges for social public safety,urban traffic supervision and urban planning.Safety accidents such as crowding and stampede often appear in crowded areas.Therefore,more and more researchers are devoted to the research of crowd counting and density estimation.The rise of deep learning has greatly promoted the pace of research on computer vision tasks.In the field of crowd counting and density estimation,the crowd counting methods based on convolutional neural network has the ability to well deal with the challenges such as non-uniform crowd distribution,scale variations,perspective distortions,light changes,etc.In this thesis,we first systematically analyzes the existing approaches of crowd counting and density estimation.And then,we presents two algorithms for tackling the problems of scale variations and perspective distortions from different perspectives.The evaluation is performed on the public datasets,i.e.,ShanghaiTech,UCF_CC₅0 and MALL,to demonstrate the feasibility and effectiveness of the proposed algorithm.The main research contents are summarized as follows:1.We introduce a scale-aware crowd counting algorithm.Its main idea is as follow.The task of crowd counting and density estimation is divided into three stages.The key role of the first stage is feature encoding,owing to the usage of the image pyramid set as input,and the diversity of multi-scale feature information expression of counting model is essentially increased.The key role of the second stage is multi-scale information extraction,and a novel scale-aware module and inverse-scale-aware module are designed to further boost the capability of counting model for extracting the multi-scale information.The main role of the third stage is to generate the crowd distribution density map.This stage is mainly composed of three crossed branches with different dilate ratios,and the feature maps obtained from different branch are exploited to generate the ultimate crowd distribution density map.In addition,the algorithm uses a skip-connection between the top convolutional layer and the bottom atrous convolutional layer to reduce the risk of vanishing gradient and gradientexplosion.The intermediate supervision strategy is used to optimize the network parameters.2.An attentive multi-stage convolutional neural network is introduced for crowd counting.Its mainly idea is as follow.The attentive multi-stage convolutional neural network is mainly composed of hierarchical density estimator,soft-attention mechanism and auxiliary counting classifier.The hierarchical density estimator adopts a hierarchical strategy to mine semantic features and multi-scale information in a coarse-to-fine manner to tackle the problem of scale changes and perspective distortions.In addition,considering that the background noise has negative influence on the quality of generated density map,a soft attention mechanism is integrated into the counting model to distinct the foreground and the background to further improve the density map quality.Furthermore,inspired by multi-task learning,an auxiliary counting classifier is integrated into the counting model to perform counting classification tasks,thereby enhancing the ability of representing semantic information.Extensive experimental results demonstrate the effectiveness and feasibility of the introduced algorithm in coping with scale changes,perspective distortions.