Crowd Density Estimation Based on Convolutional Neural Network

For tasks like crowd density estimation and crowd counting, the model needs a larger receptive field. The easiest way to obtain a large receptive field is to use a large convolution kernel, which increases the model's parameters. Therefore, to realize a large receptive field, the model parameters are not excessive and the model performance is not bad. Based on ResNet, this paper proposed a new crowd density estimation model. Two sets of long convolutions were superimposed on the residual network. While obtaining a broad receptive field, the model can have better performance and reduce the number of model parameters. The improved model was trained and tested on three datasets, ShanghaiTech, UCF-QNRF and UCF_CC_50. Compared with the MCNN model, MAE and MSE on the ShanghaiTech dataset PartA decreased by 36.7 and 45.3. They decreased by 18.7 and 28.9 on PartB. On the dataset UCF-QNRF, they decreased by 179.1 and 247.1. On the dataset UCF_CC_50, they decreased by 122.5 and 95.7. Compared with other models, the experimental results showed that the improved model has greatly improved on various performances, the accuracy of crowd number estimation has greatly increased, and the two evaluation indicators of average absolute error and mean square error have certain advantages under this network, proved the practicability, effectiveness and stability of the proposed model.