| In recent years,with the rapid development of social economy and the continuous improvement of people’s living standards,people have gathered more and more frequently in public places,which can cause a series of potential safety issues.Therefore,it is of great practical significance to estimate crowd density in public places.Crowd density estimation is an important research direction in the field of computer vision.It refers to accurately estimating the number of people and generating a crowd density map that reflects the density of people in the image.Driven by deep learning in recent years,crowd density estimation have obtained many achievements.However,there is still a problem of perspective distortion in the image,which greatly affects the accuracy of crowd density estimation.In view of the above problem,this thesis focuses on the crowd density estimation method based on deep learning.The concrete research content is as follows.Firstly,in view of the problem of perspective distortion in the field of crowd density estimation,a crowd density estimation method based on the multi-scale dilated convolution is proposed in this thesis.This method utilizes multiple dilated convolutional neural networks with different receptive fields in parallel to extract multi-scale features,making full use of the feature that the dilated convolution obtains a larger receptive field without reducing the resolution of the image.In addition,the loss function is improved and the image quality evaluation index is integrated into the loss function.Experiments prove that the method proposed in this thesis is good at improving the accuracy compared with the classical methods in recent years.Subsequently,the crowd density estimation method based on the multi-scale dilated convolution is improved and a crowd density estimation method based on adaptive receptive field is proposed in this thesis.It replaces three parallel convolutional neural networks with three independent crowd density estimation networks.This method crops the input image into patches and classifies them,so that the patches adaptively choose networks corresponding to the receptive field for crowd density estimation.Experiments show that the crowd density estimation method based on adaptive receptive field proposed in this thesis can significantly improve the accuracy compared with the crowd density estimation method based on multi-scale dilated convolution and the classical methods in recent years. |
