To avoid the interference of depth of field and occlusion and improve the accuracy of crowd counting, three models of LeNet-5, AlexNet and VGG-16 were adopted and the characteristics of objects with different depth of field in the image were extracted. After adjusting the convolution core size and network structure of the above model, model fusion was carried out. A deep convolution neural network structure based on multi-model fusion was constructed. In the last two layers of the network, the convolution layer with convolution core size of 1×1 was used to replace the traditional full connection layer. The extracted feature maps were integrated with information and the density maps were output. The network parameters were greatly reduced and some improved data was obtained. The efficiency and accuracy of the algorithm were taken into account. The theoretical analysis and experimental verification were carried out. The results show that, in public population counting data set of two subsets of shanghaitech and UCF_CC_50, the mean absolute error and mean square error of this method are 97.99 and 158.02, 23.36 and 41.86, 354.27 and 491.68, respectively. It achieves better performance than the existing traditional crowd counting methods. At the same time, migration experiments are carried out. It proves that the population counting model proposed in this paper has good generalization ability. This study is helpful to improve the accuracy of population counting.