Enhanced crowd counting with weighted attention network and multi-scale feature integration

Abstract

Crowd counting plays a crucial role in the field of computer vision, particularly in practical applications such as traffic monitoring. However, current methods that establish mappings between original images and density maps are not only prone to overfitting but also struggle with occlusion and scale variation in crowded scenes. In this paper, we propose a novel Weighted Attention Focusing Network (WAFNet) to enhance crowd counting performance by decoupling the image-density mapping. Our approach first employs a two-stage model to separate the image density map. It then introduces a weight map, generated by the front-end network, to address the issue of scale variation. Additionally, we incorporate a Multi-Layer Feature Compilation Module (MLFCM) to better preserve and fuse features from multiple layers and adopt a Low-Resolution Feature Enhancement Module (LRFEM) to enhance the low-resolution features of the crowd. Experiments conducted on six benchmark crowd counting datasets demonstrate that our method achieves improved performance, particularly in dense and occluded scenes.