Research On Crowd Counting Algorithm Based On Deep Learning

In the public safety protection industry,crowd counting and crowd density estimation play an important role.Through the estimation of crowd density in public places,timely and accurate early warning of public safety accidents,systematic and comprehensive guidance of public safety accidents,and full and effective post-event control of public safety accidents can be realized.At the same time,the widespread application of deep learning in the study of population statistics makes public safety management more convenient and quicker.Therefore,the study of crowd counting based on deep learning is widely favored by academic researchers.At present,most crowd counting models use convolutional neural networks as the basic framework.By inputting crowd images or video surveillance information in public places into the crowd counting network for crowd density estimation,comprehensive personnel positioning and number statistics can be realized.However,the research on the crowd counting model still faces many difficult problems that need to be solved,such as the complex background information in the image data,the serious occlusion of the object,the uneven distribution of people,and the dense crowd.In order to effectively solve the above difficult problems and enhance the counting performance of the crowd counting network model,this paper has carried out a series of research based on convolutional neural network technology.The main innovations are as follows:(1)In order to solve the problems of complex background,occlusion between people or between people and objects,and uneven distribution of crowds in the process of counting people,a crowd counting algorithm integrating dual attention mechanism is proposed.By improving the common crowd counting network CSRNet,a joint lossbased space-channel dual attention network(JL-SCDANet)is proposed,which adopts joint loss function and introduces spatial and channel dual attention mechanism.The structure of the model mainly draws on the three-stage design method.The front end and back end of the model keep the original network structure of CSRNet unchanged.The middle end is connected in series with plug-and-play spatial attention module and channel attention module to focus on the position of the image crowd and emphasize the important information of the head.The joint loss function is used in the training process of the model,and the optimal hyperparameter values are adjusted through a large number of experiments.This model is mainly trained and tested in three common public datasets Shanghai Tech Part B,Mall and UCF＿CC＿50.The test results show that JL-SCDANet can effectively improve the accuracy of people counting,and has good generalization ability and robustness.(2)In order to solve the problems of uneven distribution of people,inconsistent size of head targets,and difficulty in counting heads when the crowd is too dense,further research is carried out on the basis of the crowd counting algorithm that integrates the dual attention mechanism,and a population statistics algorithm that integrates dual-line features is proposed.The Double Line Feature Network Model Combining Multiscale Feature Fusion and Attention Mechanism(DLFNet)is established by improving the proposed JL-SCDANet network architecture.In the DLFNet model,the image first passes through the coarse-grained feature extraction module,and then enters the multiscale attention module and the CBAM attention module respectively to achieve bilinear feature extraction.The multi-scale attention module is used to realize the feature extraction work with dense crowds or small personnel targets in the picture.CBAM is used to accurately capture the key personnel information of the image and filter the unimportant background information.After the fusion of the features extracted by the two modules,the attention module is used again to realize the re-calibration of the feature information and filter the impurities of the fusion features.Finally,it enters the dilated convolution module to extract the deep two-dimensional features and output the predicted crowd density map.A large number of comparative experiments show that the DLFNet model can still effectively extract the head feature information of small targets when the crowd distribution is dense,reduce the statistical error of the number of people,and improve the accuracy of the number of people.In summary,based on the convolutional neural network technology,this paper proposes two population statistics models JL-SCDANet and DLFNet,and conducts comparative experiments on multiple public data sets.The experimental results verify that both models have achieved low error values in counting accuracy and have high practical application value.