Crowd Counting Based On Convolutional Neural Network

With the continuous development of deep learning,Convolutional Neural Networks have been widely applied to various computer vision fields.For the problem of crowd counting in image,this paper mainly uses convolutional neural network to solve the crowd counting error caused by the different scale,uneven distribution,complex background and so on.The specific research contents are as follows:1.Crowd counting by the global-aware network with consistency loss constraints.The density of crowd plays a key role in generating accurate density maps.Based on this starting point,this paper proposes a global-aware network,which consists of two parts: the first part uses the first ten layers of the VGG network as our basic backbone network;The second part of the network consists of feature extraction branche and global-aware branche.This design adds a global-aware branch after the feature extraction branche,uses the global pooling layer to obtain the mean output of the feature map,scales the output of the global pooling layer to the size of the input feature map.The output features of the global-aware branch and the features of the extraction branch are fused together to generate a density map.Based on the phenomenon that the number of persons included in the original image is equal to the sum of the number of persons included in the cropped subgraphs,the consistency loss function is proposed.The consistency loss is combined with Euclidean loss as the final loss function of the network.2.Crowd counting by the dual-branch scale-aware network with ranking loss constraints.Crowd images captured from different views result in scale variations of heads,this paper proposes a dual-branch scale-aware network,which consists of two main components: the first part uses the first ten layers of the VGG network as our basic backbone network;and a dualbranch(Branch_S and Branch_D)network is proposed to be the second part of the network.Branch_S extracts low-level information(head blob)through a shallow fully convolutional,and Branch_D extracts high level context features(faces and body).Features learnt from the two different branches can handle the problem of scale variation due to perspective effects and image size differences.Features of different scales extracted from the two branches are fused to generate predicted density map.On the basis of the fact that an original graph must contain not less than the number of persons than any of its sub-images,a ranking loss function utilising the constraint relationship inside an image is proposed,the ranking loss is combined with Euclidean loss as the final loss function.3.Crowd counting by the multi-channel fusion group convolutional neural network with structural similarity index loss function.The variability of head size and the diversity of crowd distribution are the two main challenges in image crowd counting.Many methods try to solve these problems by using multi-column or multi-branch networks,but due to the limitation of the number of columns or branches,the scale of extracted features is limited.This paper proposes a multi-channel fusion group convolutional neural network for dense crowd counting,the network is composed of two major components: the first part uses the first ten layers of the VGG network as our basic backbone network;The multi-channel fusion group convolutional module is proposed to be the second part of network,the multi-channel fusion group convolutional module is the key components of this network.Each group convolution module in this module is densely connected with other layers to obtain different levels of features,meanwhile,we introduce group convolution to reduce network parameters.In this paper,a density map is generated to complete the counting task,and a structural similarity index loss function is proposed based on the structural similarity index of the images,the structural similarity index loss function is combined with the Euclidean distance loss function as the final network loss function.This paper proposes three crowd counting algorithms from three different or the same angles,and compares with the existing algorithms on three public data sets,and obtains competitive experimental results...