Design And Implementation Of Crowd Counting Network For Complex Congested Scene

With the continuous updating of science and technology,modern intelligent systems are more and more widely used in daily life,and video surveillance devices are also everywhere.As an important task in video surveillance systems,crowd counting has attracted more and more researchers' attention.Accurately estimating the number of people in video images is of considerable significance to public social safety,and it is a great challenge to count people in complex and dense scenes accurately.Crowd counting has many problems in practical application.For example,external factors such as obscured target,cluttered background,uneven illumination,and significant differences in human head scales make it difficult to estimate the number of people accurately.Besides,the crowd counting algorithm based on the convolutional neural network,the convolution operation only models the local area,and the global information is lost,which makes the extraction of context information difficult,and also affects the accuracy of crowd counting.Therefore,this paper focuses on the challenges of background clutter,scale change,and the limitations of convolution operations,and studies the crowd counting network in complex dense scenes.To solve the problem of changing the scale of the crowd picture,this paper designs a crowd counting network fused with dilated convolution.The dilated convolutional layer expands the receptive field range and extracts multi-scale crowd feature information,thereby improving the accuracy of the output density map of the crowd counting network.To solve the problem of cluttered background in crowd pictures,this paper designs a crowd counting network that integrates attention mechanism.The use of the attention module has designed two types of networks in series and parallel.By adding the attention module,the network can focus on the head position of the crowd to a greater extent and filter out the background information,thereby improving the robustness of the crowd counting network.On this basis,the dilated convolution and attention mechanism are jointly used in the same counting network,and a multi-branch network structure is designed.The above four networks all use the Shanghai Tech dataset Part?A for experimental verification.The experimental results show the effectiveness of the hollow convolution and attention mechanism for crowd counting networks.In view of the limitation of convolution operation,this paper designs two different crowd counting network structures to compensate for the loss of global information and promote the extraction of context information.The first is a crowd counting network based on self-attention knowledge distillation.By using the self-attention distillation strategy at different layers of the network to guide the refinement of the features of the shallow network,it further promotes the refinement of context information by the deep network,making the density map generated by the network more similar to the true density map.The second is a crowd-counting network structure based on global inference.By adding a global inference unit to the network to model the association of long-distance regions on the feature map,increase global features and improve the accuracy of the crowd counting network.Both networks use the UCF?QNRF data set for experimental verification.The experimental results show that self-attention knowledge distillation and global inference both reduce counting errors and greatly improve network performance.