Research On Crowd Counting Method Based On Convolutional Neural Network

With the continuous growth of the world’s population and the frequent occurrence of trampling incidents in various countries,crowd counting has played an increasingly important role in areas such as intelligent monitoring and public environmental safety.However,due to factors such as complex background and crowd occlusion,the accuracy of crowd counting cannot meet the actual needs.Therefore,how to obtain better accuracy of crowd counting has become an urgent problem in the field of crowd counting.Based on a comprehensive analysis of the current research situation at home and abroad,combined with the advantages and disadvantages of existing crowd counting methods,and using deep learning and other relevant knowledge,we conduct in-depth research on how to obtain better crowd counting results.Firstly,analyze the relevant methods and research status of crowd counting,and briefly introduce the basic layers of deep learning and convolutional neural networks and some classic convolutional neural network structures.Secondly,we propose a crowd counting method based on cross-resolution feature fusion.This method firstly designs a multi-resolution network,which connects the low-resolution sub-network to the high-resolution basic network.It keeps the high-resolution feature map always exist,and reduces spatial information loss when feature maps of different resolutions are transformed between each other.Secondly,the method fuses feature maps of different resolutions to obtain multi-scale feature information.Finally,the features are input to a deconvolution network to obtain the crowd density map,and the population quantity is obtained by integrating the density map.Thirdly,a method for crowd counting based on cross-fusion efficient convolutional neural network is designed.This method firstly obtains a new feature map containing more semantic information by cross-fusion of the features of different convolutional layers;secondly,when the convolution operation is performed according to the traditional convolution kernel,the receptive field has a single characteristic,and an efficient convolution kernel is introduced.On the basis of the convolution kernel,the sub-assisted convolution kernel is split to obtain feature maps from different receptive fields to improve the accuracy of the crowd counting model.Finally,the image to be tested is input into the model with improved accuracy to obtain the crowd counting result.Finally,the two crowd counting methods proposed are verified on the Shanghaitech and Mall datasets,proving the feasibility,effectiveness and generalization of the algorithm.