Research And Application Of Crowd Counting Methods Based On Two Models

Nowadays,with all-round rise of theories in the fields of artificial intelligence,and great development of computer vision technology and digital image processing technology,it promotes the development of crowd counting and density estimation,and crowd counting is already everywhere in our life,such as classrooms of schools,entrances to malls,monitoring sites of streets and monitoring center of transportation squares.Although the object-detection approach with people's heads or head-shoulder parts detects and locates the crowds to get its amount,it performs poorly at the outdoor scene of crowdedness and great numbers of people.Aiming at outdoor scenes,the two goals of density estimation are generating high-quality density maps and getting and the total population,which is full of great research significance and application value,such as crowd monitoring and human traffic management.Based on crowd density estimation,in this thesis,it focuses on outdoor scenes with massive human traffic,such as squares,streets and stations.It also aims at figuring out some problems about low quality of generated density maps and inaccurate counting results of crowds.With above goals,we carry out research and design work in the field of crowd counting methods,design and develop an outdoor-scene-crowd-counting system.This thesis proposes a novel encoder-decoder network,called SED-CNN,which applys the concepts of encoder and decoder to its end-to-end architecture.In the encoder,we use multi-column multi-scale blocks,which extracts feature details in a given image effectively.About our decoder,it's a single-path decoder with a set of different size kernels of deconvolutions.In order to optimize the training process of our network,we use a jointly final loss combining MSE loss with SSIM loss and Relative Count loss to train our model,obtaining better results than only using MSE loss.What's more,we conduct extensive experiments on three benchmark datasets and it achieves the best overall performance,which is a significant demonstration of the effectiveness of this approach.Besides,it proposes a method focused on crowd density estimation,which is called CDEcGAN and based on conditional generative adversarial network.Inspired by the solution of image-to-image problem through a conditional generative adversarial network,we input crowd images and their ground-truth density maps as extra conditional information to optimize the training process.To be specific,its generator updates the network based on SED-CNN,and its discriminator is designed based on PatchGAN,which is an algorithm of patch discrimination and focuses on the high-frequency parts of a generated image to improve its quality.The experiments demonstrate that the proposed model is practicable and effective.Moreover,this thesis designs and develops a system of outside crowd density estimation with above two methods on different architecture.