In this paper,we propose a population counting model of deep adversar y network,which is mainly divided into generator network and discriminator network.First,the first ten layers of VGG16with good migration ability and feature extraction ability are used as the front-end modules to initially extract features;then,in response to th e serious occlusion and uneven distribution of the crowd,the deep dilated convolution module designed by us is used to aggregate the crowd in formation,and the shallow and deep head features are fused to enhance the adaptability of the network to the crowd.In this process,th e expansion convolution is used instead of the traditional convolution layer to extract the features of the image without losing the resolu tion of the image.Finally,the density map and the label density map are input into the discriminator network to distinguish,so as to ge nerate the density map more similar to the label density map and improve the accuracy of population counting.The experimental results show t hat,compared with other methods,the method in this paper has better results in both objective indicators and supervisor vision.