Abstract: Aiming at the problems of scale variation, uneven distribution, and background occlusion of dense crowds in complex scenes, a crowd counting algorithm MFFBSNet based on multi-scale feature fusion and background suppression is proposed.The first 13 layers of the visual geometry group network VGG-16 are utilized as the front-end of the network. An atrous spatial pyramid pooling (ASPP) and a pyramid split attention (PSA) mechanism based on a lightweight design are introduced to construct a multi-scale feature fusion module, which addresses the problem of scale variation in dense crowds； In the middle of this network, spatial and channel attention mechanisms are incorporated to refine the feature maps, highlighting the head regions in the image;  The backend of this network employs atrous convolution, which enlarges the receptive field without losing image resolution, to generate a background segmentation attention map. This suppresses background noise in the image and enhances the quality of the crowd density map. Experimental results on three public datasets, namely ShanghaiTech, UCF_CC_50, and NWPU-Crowd,demonstrate that the proposed crowd counting algorithm based on the MFFBSNet achieves higher counting accuracy compared to methods such as MCNN,SwitchCNN,and CSRNet.

Key words: dense crowd counting, multi-scale fusion, background suppression, density map