Spatial Joint Upsampling Based Crowd Counting

Crowd counting are playing an important role in controlling public crowd flow, maintaining public safety order, and controlling novel coronavirus (2019-nCoV). In recent years, researchers have proposed many excellent crowd counting methods, but these methods still have some problems such as insufficient information obtained in the output density map. Based on it, this paper proposed a crowd counting method based on Spatial Joint Upsampling(SJU). The method uses VGG-16 as the backbone of front-end. Meanwhile the back-end network extracts rich pixel details and spatial context information by adding cascaded joint upsampling model to combine multi-layer high-resolution feature maps, which ultimately have reached the effect of reducing the computational complexity and lifting the accuracy. Experimental results show that SJU has better performance than most crowd counting methods. This method has achieved MAE/MSE of 62.1/99.8 and 7.6/11.5 on the two parts of the public dataset ShanghaiTech A and B, respectively. The MAE/MSE after conducting curriculum learning on this dataset is 61.3/99.2. In addition, this method has also achieved excellent results on the large-scale complex crowd dataset NWPU-Crowd, with a MAE/MSE of 105.1/419.3, indicating that the Spatial Joint Upsampling(SJU) network has outstanding performance in the task of crowd counting in complex scenes.