Multi-scale Convolutional Neural Network Models For Crowd Count Estimation

With the increase of world population,dense crowds may produce great threats to the social public security and bring a lot of challenges to the urban planning.Counting the number of people in dense crowds,i.e.crowd counting,has a very good application prospect in the practices.For example,conducting effective crowd control in crowded places,such as shopping malls,stations,tourist attractions and sports grounds,can prevent the occurrence of stampedes.Methods for crowd counting can also be applied to the counting of other objects,such as vehicles on the highway,bacteria or cells under microscopes.Therefore,crowd counting has aroused great interest and attention among researchers.At present,traditional crowd counting methods mainly include three categories:detection-based methods,regression-based methods and methods based on convolutional neural network(CNN).Among them,the detection-based methods are sensitive to severe occlusions and clutter backgrounds,while the regression-based methods cannot solve the scale and perspective variations well.This thesis mainly focuses on the CNN-based methods,and proposes three kinds of crowd counting models based on the scale-adaptive convolutional neural network(SaCNN).We hightlight the work of this thesis as follows:(1)A crowd counting algorithm based on residual multi-scale convolutional neural network(RMsCNN)is proposed.In order to relieve the issues of gradient diffusion,gradient explosion and network degradation existing in SaCNN,this thesis proposes RMsCNN,which introduces the idea of residual and adds a residual learning unit between two convolutional layers.Experiments on public datasets verify the effectiveness of RMsCNN.(2)A crowd counting algorithm based on multi-column multi-scale convolutional neural network(McMsCNN)is proposed.In order to further sovle the issues of severe occlusions and the scale and perspective variations that exist in crowd images commonly,this thesis introduces the idea of multi-column model and proposes McMsCNN.The novel algorithm has three parallel subnetworks with the same structure,and each subnetwork uses convolution kernels of different sizes to correspond to different sizes of heads.Experiments on public datasets verify the effectiveness of McMsCNN.(3)A crowd counting algorithm based on two-task multi-scale convolutional neural network(T2MsCNN)is proposed.In order to improve the adaptability and sensitivity of SaCNN to the images with different dense degrees,this thesis introduces a dense degree classifier based on SaCNN and presents T2MsCNN.The dense degree classifier can classify the images according to their dense degree.Thus,the whole network can learn the change of dense degree.Experiments on public datasets verify the effectiveness of T2MsCNN.