Zooming Fusion Network: A Multi-scale Approach for Precise Crowd Counting and Individual Localization in Low-Angle Scenarios

Abstract

Crowd counting is pivotal in critical domains like security surveillance, urban planning, and traffic management. It is typically accomplished byaccurately detecting the number of individuals in a scene, thereby offering essential early warning and preventative mechanisms. Currently, the widely adopted method in the industry for dense crowd counting is based on density maps. However, this method relies on the use of Gaussian kernels to create intermediate representations (density maps), making it susceptible to the accumulation of background noise and loss of density. Moreover, it fails to provide precise individual location information, which hinders downstream tasks such as crowd movement trajectory analysis and abnormal behavior detection. To overcome these limitations, we propose a point-based framework named the Zooming Fusion Network (ZFNet), which integrates multi-scale inputs with an attention mechanism module to enhance the model's adaptability to complex scenes and its focus accuracy on crowd regions.. To validate the effectiveness of the proposed method, we conduct experiments on two public datasets. The results demonstrate that ZFNet achieves leading performance, compared to other existing state-of-the-art crowd counting methods. Furthermore, to address the lack of low-angle scenes in public datasets, we introduce a LIB_CC_500 dataset, specifically tailored for low-angle crowd counting. LIB_CC_500 The model's performance is validated on this new dataset, with results indicating its good generalization capabilities.