Research On Crowd Density Estimation Based On Convolutional Neural Network

Today,the development of the national economy has led to the progress of socialist urbanization,more and more people are flocking to cities.Accordingly,the social problems brought about by it are also increasing,for example: the occurrence of unsafe accidents such as trampling caused by crowding;the pressure on the traffic dispatch caused by the increase in the waiting hall staff.To solve the above problems,it is necessary to accurately predict the number and distribution of people in the actual scene,and the image can clearly and intuitively reflect the change of the crowd in the actual scene,so the population density estimation based on static image information has important research value.In recent years,the Convolutional Neural Network model has developed rapidly in the fields of semantic segmentation,target detection and recognition due to its ability to acquire deep-level features.Researchers have also applied it to the field of crowd counting and achieved Has achieved better results,but there are still some challenges that have not yet been overcome.Based on sufficient research and analysis of the current population density estimation algorithm,this paper has done the following work:(1)Based on the analysis of the existing crowd density estimation algorithm based on convolutional neural network,a multi-column structure crowd density estimation algorithm is implemented.Multi-scale features are extracted through multiple columns of different-size convolution kernels to cope with the different sizes of head information appearing in the image,to solve the problem that single-column structure is difficult to deal with scale changes;the end of the network uses a convolution layer to replace the original fully connected layer,making the input The size of the picture is not limited,and the application range of the network model is more extensive.Theexperimental results show that the multi-column structure has certain advantages in the crowd density estimation task.(2)Aiming at the problems of multi-scale feature information loss,poor fusion and low quality of density map in the crowd counting method based on multi-column convolutional neural network.In this paper,a new crowd counting method is proposed based on encoding-decoding multi-scale convolutional neural network.The decoder part adopts multi-column convolution to capture multi-scale features,expands the receptive field and reduces the amount of calculation through the atrous convolution and space pyramid pooling,retains the multi-scale feature and the context information of the image;the decoder up samples the encoder output,Realize the effective fusion of high-level semantic information and low-level feature information at the front end of the encoder,thus fully ensuring the effective use of feature information at all levels.Experimental results show that the structure not only improves the accuracy of crowd density estimation,but also reduces the computational complexity to a certain extent.(3)Aiming at the problem of average pixel fusion of multi-channel information in the multi-column convolutional neural network crowd density estimation algorithm resulting in pixel loss in the output density map,the Conditional Adversarial Generation Network is applied to the multi-column convolutional neural In the network,a crowd density estimation method based on multi-scale conditional generation neural network is proposed.Two network models,generator and discriminator,are established.The generator model uses a multi-column convolutional neural network to extract multi-scale features;the discriminator network uses a five-layer convolutional structure to output similarity probabilities.The prediction density map closest to the original truth map is obtained through the game of the two networks,and the multi-scale information is aggregated together in a cooperative manner through a special loss function,and finally the prediction density map close to the true value is obtained infinitely.Experimental results show that the population density estimation result of this method is more accurate,and a high-quality predicted density map can be generated.(4)Drawing on the advantages of the multi-scale convolutional neural network crowd density estimation method of the encoding-decoding architecture and themulti-scale conditional anti-generation neural network crowd density estimation method,the above two methods are combined together for the actual scene needs,and proposed Encoding-Decoding Multi-Scale Convolutional Neural Network(EDMSCNN)is proposed based on multi-scale conditional anti-generation neural network crowd density estimation.EDMSCNN is used as the generator network model to extract multi-scale features,and the generator uses a five-layer convolution structure for cooperative fusion of multi-scale features.Under the condition of ensuring sufficient extraction of multi-scale features,the multi-scale features are aggregated together in a cooperative manner through the adversarial network idea.The experimental results show that the population density estimation accuracy of this method is optimal.