Research on Adaptive Density Normalized Crowd Counting Algorithm

Aiming at the problems of uneven distribution of crowds in static crowd images, changes in head size, and diverse crowd densities, the accuracy of crowd counting decreases. This paper proposes an adaptive density-normalized crowd counting network YOLO-FPDM-DNM., the network improves the Darknet-53 structure, improves the feature extraction ability, introduces adaptive geometric propagation parameters in the Gaussian density module to solve the problem of head size change, and designs an adaptive density normalization module to solve the problem of excessive crowd density in some areas. The resulting network training is difficult and the accuracy rate is reduced. The experimental results show that compared with the DM-Count network structure with the best effect in the UCFQNRF dataset in recent years, the MAE is reduced by 2.9. MSE decreased by 4.7.