An Improved HRNET and its application in crowd counting

Abstract

Aiming at the low accuracy of crowd counting caused by scale change and occlusion in dense scenes, this paper proposes to generate the truth map into non overlapping independent areas in HRNet to facilitate the crowd location statistics of network density map; Then the 3D attention mechanism is introduced to make the network focus on the useful information of the feature map; Finally, during the training, the mean square error loss (MSE loss), L1 loss and cross entropy loss are combined into the total loss function to optimize the generalization ability of the model; The combination of the above methods improves the accuracy of the model in crowd counting and crowd location. Compared with the main methods in recent years in the public datasets NWPU, Shanghai Tech, the experimental results show that the proposed model can effectively improve the accuracy and robustness of crowd location counting.