Research On Crowd Couunting Based On Convolutional Neural Network

As the development of globalization,the frequency of group activities and behaviors is getting higher and higher.As the effective tool to control crowd flow and analyze the process information for crowds,intelligent monitoring is widely used in a variety of places.Further,if we want to analyze the crowd activities,the most important and fundamental indicator is crowd density.Recently,the study of crowd counting has undergone a great leap forward,but facing complex and changeable crowd scene,there are still a number of challenges.For the past few years,convolutional neural network(CNN)has become a hot topic in the field of computer vision.Because of its excellent ability for feature extraction and generalization,CNN is widely used in different types of computer vision tasks.Similarly,CNN can extract the counting features of crowd effectively in complex background.Whereas,current deep counting network often implement multi-columns or multi-scales structures,which cause that the counting performance is severely constrained,due to a series of the problems existing such as nuisance parameter and difficult training.In order to solve problem of scale-dependent feature extraction,based on VGG network and deep residual theory,this paper designs a residual building block convolutional neural network(RBBCNN).Based residual building block as the basic unit,the entire network is divided into two stages:encoding and decoding.For the different in-out relations,three different residual building blocks are embedded to achieve concrete residual connection.In short,the RBBCNN consider deepening the network to improve representational capacity of the model and facilitating training both ways.Furthermore,for verifying the validity of residual connection,the validity analysis of contrast test is carried out.As the experiment on three data sets ShanghaiTech,UCSD and MALL with different density,RBBCNN achieves lower counting error in both of sparse and dense crowd scenes.Moreover,this paper also designs a deep counting model based on cross-level parallel way called CLPNet.The CLPNet fuses cross-level and multi-scale features effectively to solve changing head scale.To adapt to the fusion,five scale aggregation modules are designed to fuse different level semantic features.Another measure for solving changing density and scale is that the network output two-channel density map to make better area matching between head and default Gaussian kernel,considering reducing counting error from the global picture level.Further,to make up for the shortfall of Euclidean loss,the structural similarity loss is introduced to measure correlation of surrounding pixel and the above losses jointly optimize the network parameters.At the particular data set,we also perform comparative analysis of validity.At the same time,we conduct comparative test and result analysis on the four mainstream data sets,showing that CLPNet has higher counting accuracy and good generalization capability.