Crowd Density Estimation Method Based On Convolutional Neural Network

With the continuous progress of modernization and urbanization,large-scale crowd mass has become more common,which brings huge hidden dangers to the public safety.Crowd counting plays a quite important role in the field of intelligent video surveillance.It can estimate the total number of people from an image and provide real-time warnings,which can effectively avoid safety accidents.With deep learning technology,especially computer vision,the crowd density estimation method based on the convolutional neural network can not only reflect the spatial distribution of the crowd,but also gains high counting accuracy.Therefore,it has attracted wide attention with extremely high practical value.However,there are still some severe challenges in crowd counting,such as the scene transition,the background interference,and the real-time requirements in the real scenario,which restrict the performance and application of the counting model.To address these problems,we conduct research on density map generation,feature fusion,network compression in this thesis,and propose targeted solutions to further improve the counting performance and the practical value of the counting model.The main contributions and innovations in this thesis are as follows:1.To address the contradiction between the resolution and counting accuracy of the density estimation map generated by the counting network,an elaborate density estimation algorithm is proposed.This algorithm can enhance the semantic representation ability of low-level features by developing a multi-scale feature fusion module,which is a top-down architecture with lateral connections.Then,a density regression networkbased on multi-task learning is designed to solve the background interference problem by adopting auxiliary tasks of semantic segmentation.Next,the proposed method introduces the coarse-grained density representation into the high-resolution density estimation map through the adaptive density fusion module.Finally,to avoid the vanishing gradient,the combined loss function based on distributed supervision is applied to refine the density estimation map layer by layer.Extensive experiments on multiple datasets show that the proposed algorithm can not only effectively reduce counting errors,but also generate high-resolution density estimation maps.2.To address the problem of spatial misalignment and semantic inconsistency between adjacent features in the fusion stage,we propose a density estimation algorithm based on coordinated feature fusion.The algorithm adopts a strong baseline network and two embeddable convolution modules,including the spatial alignment module and the semantic consistency module.The spatial alignment module can learn the pixel offset transformation relationship of the features to alleviate the spatial dislocation caused by the feature resolution difference.The semantic consistency module applies the multiscale attention mechanism to capture the pixel-by-pixel weight,which can mitigate the semantic inconsistency caused by the feature level gap.Extensive performance evaluation and ablation studies show that the proposed algorithm can enhance the effectiveness of feature fusion and further improve the counting performance and generalization ability of the network.3.To address the problems of huge model parameters and long inference latency of counting networks,a density estimation algorithm with a lightweight backbone is proposed.First,the algorithm adopts the lightweight backbone to extract multi-scale features.Then,a parameter-sharing context-aware module is constructed to capture global context information without increasing the number of parameters significantly.Finally,the weight-sharing mechanism can realize the automatic switching between multiple density generation modules in the training stage to reduce the amount of floating-point operations.Extensive experiments of performance and parameter comparison show that the proposed algorithm can achieve the optimal trade-off between counting accuracy and computation efficiency.