Crowd Counting Based On Convolutional Neural Network

With the continuous advancement of economic globalization and the continuous improvement of China’s urbanization rate,crowd counting agglomeration is becoming increasingly apparent.However,crowds are not only prone to stampede accidents when gathering activities but also easily lead to clustered epidemics when COVID-19 is not over yet.In this case,it is difficult to detect and warn the agglomeration of people in time and make an accurate judgment only by manual treatment.Therefore,it is essential to make accurate population statistics to prevent accidents.Thanks to the emergence of artificial intelligence algorithms with deep learning and convolutional neural network as the core,the fields of image processing such as target detection,recognition,and segmentation have developed rapidly,and the crowd counting algorithm’s accuracy has been dramatically reduced improved.However,the crowd counting algorithm still faces the following problems: complex background of crowd picture,significant variation of head scale,uneven distribution of crowd,and severe occlusion.We propose a new parallel cross-layer network to solve the above problems,which takes VGG as the backbone and introduces the SSIM loss function with a cross-layer multi-scale fusion module.The specific work contents are as follows:(1).A parallel cross-layer feature fusion crowd counting network algorithm is proposed.VGG16 is used as the feature extraction backbone to extract the bottom features of the picture.At the same time,to solve the problems of population-scale differences and uneven population distribution,this method uses two networks with different scale feature extraction capabilities to extract the bottom features output by VGG16.The Dense branch uses a convolution layer with smaller convolution kernels to extract smaller-scale details,while the sparse branch uses dilated convolution and convolution layer with larger convolution kernels to extract larger-scale details.(2).Aiming at the problem of inaccurate counting caused by significant scale differences and complex picture background,a cross-layer multi-scale fusion module is designed in this paper.In order to use the semantic details of the bottom feature and the high-level feature,and the features with different scales at the same time,this module can fuse the bottom feature directly from VGG16 with more location details and the high-level feature with more semantic details after more feature extraction convolution layers,and fuse the features with different scales at the same time.(3).A larger convolution kernel and pooling layer are usually used to extract features on a larger scale.In this paper,many dilated convolutions are used to replace the large convolution kernel,which improves the receptive field but does not like pooling,which is easy to lose some details.It not only ensures the image resolution and reduces the training parameters but also improves the accuracy of crowd counting in dense scenes.(4).To improve the authenticity of the density map generated by the crowd counting network,this training method introduces the SSIM(Structural SIMilarity)and generates an adaptive loss function combined with the European loss function.The experimental results show that this loss function can help the model achieve better results than others.Combined with the above methods,we propose a multi-layer parallel convolutional neural network based on cross-layer fusion and multi-scale aggregation,which is employed in these public datasets: Shanghai Tech,UCF＿CC＿50,UCF＿QNRF and Shanghai EXPO’10.The results show that the crowd counting algorithm proposed in this paper has higher accuracy and more vital generalization ability than the existing methods.