| Crowd density estimation aims to estimate the actual number of people in images with related algorithms,which is of great significance for maintaining public safety.Limited by scale variation and complex background,crowd density estimation is still a challenging computer vision task.The existing deep learning-based methods adopt a multi-channel network and have stronger multi-scale feature extraction capabilities.However,each branch is independent in the multi-channel structure of existing methods,and the transfer of features and the extraction of attention information are ignored.The performance of the algorithms is severely limited.In view of the above problems,the research of this paper are as follows:(1)A crowd density estimation network based on multi-channel feature fusion is proposed.Firstly,based on the problem that the existing methods lack information transmission between channels,the feature fusion module is designed to construct the information transfer path between channels.Dilated convolutions with different dilation rates are employed by this module to extract multi-scale features,which improves the aggregation ability of multi-scale feature of the network.Secondly,the density map regression module is designed to reduce the loss of multi-scale features.(2)A crowd density estimation network based on attention and feature fusion is proposed.Based on the previous work,attention mechanism is introduced to reduce the interference of complex backgrounds.Firstly,the channel attention module is designed to capture both local and global attention information.Secondly,the spatial attention module is designed to generate attention maps that enhance activation responses to crowd-regions.(3)A crowd density estimation network based on pyramid feature fusion is proposed.In order to solve the over fusion of features in the feature fusion module of multi-channel feature fusion network,a pyramid feature fusion module is proposed.Bottom-up and upbottom feature fusion paths are constructed to aggregate multi-scale features.Compared with previous methods,this method has a neater framework and higher accuracy.In this paper,sufficient experiments are carried out on Shanghai Tech dataset and UCF＿CC＿50 dataset,and the three crowd density estimation methods in this paper are compared with the current advanced crowd density estimation algorithms,which proves the effectiveness and robustness of the three methods proposed in this paper. |
