Research On Crowd Counting Method Based On Convolutional Neural Network

Crowd counting is a hot research topic in the field of computer vision,playing an important role in video surveillance,public safety,and crowd management.However,the current methods for crowd counting still face significant challenges,as the multi-scale variation of targets and irrelevant background interference can affect the accuracy of the counting network.Existing convolutional neural network-based crowd counting algorithms still suffer from problems such as partial information loss and low counting accuracy.In response to these challenges,the main research work of this manuscript is as follows:(1)Design a crowd counting network,called CAENet,using density map regression to resist multi-scale variation and background interference.The network uses the encoder-decoder structure as the backbone,and the encoder uses the feature extraction network of VGG-16.In order to deal with the multi-scale problem,the multi-scale fusion module of the network is designed to fuse the features with different scale semantic information in the encoder.The attention module is designed with a separate decoding channel,and the attention maps generated by the module are fed to each stage of the decoder to suppress background interference.The network completes the training step by step in a supervised manner and uses the output density map of the last layer as the final prediction.(2)The focal inverse distance transform is used to generate the FIDT map served as the label,and a crowd counting algorithm based on improved HRNet is proposed utilizing FIDT map regression.The algorithm uses HRNet as a global feature extractor and replaces upsampling and downsampling with DUC modules and Passthrough layers in the original network of HRNet,respectively,to reduce local information loss in the feature extraction process.It constructs a feature fusion module at the end output of the network to further weaken the impact of irrelevant background and noise while fusing features at different resolutions.The model is trained by using the FIDT map regression and outputs the final prediction result.The two counting methods proposed are compared with representative crowd counting models in recent years on the Shanghai Tech,UCF＿CC＿50,and NWPU-Crowd datasets to verify their accuracy and robustness.At the same time,an ablation experiment is designed to verify the effectiveness of each module proposed in this manuscript on the network.