Wednesday 29 October 2025
A foggy day is a challenge for crowd counting algorithms, which struggle to accurately detect and count people in hazy conditions. But researchers have now developed a new approach that combines physical principles of atmospheric scattering with deep learning techniques to achieve improved results.
The problem of crowd counting in foggy environments arises from the degradation of visual features caused by long-range target blurring, local feature degradation, and image contrast attenuation. Traditional methods rely on hand-designed features or geometric perspective models, but these often fail under inclement weather conditions.
To tackle this issue, scientists have developed a new framework that incorporates physical priors of atmospheric scattering into deep learning architectures. The approach is based on the concept of transmittance dynamic estimation and scattering parameter adaptive calibration, which accurately quantifies the nonlinear attenuation laws of haze on targets with different depths of field.
The framework consists of three core components: a Physics-guided module, an MSA-KAN (multi-scale attention Kolmogorov-arnold network) layer, and a Weather-dynamics-aware GCN (graph convolutional network). The Physics-guided module employs physical parameter estimation networks to model fog attenuation effects at the pixel level. This allows for accurate quantification of haze-induced feature degradation.
The MSA-KAN layer is designed to extract features from foggy images while incorporating physical prior knowledge through parametric basis functions. This enables the model to effectively discern degraded features and reduce feature confusion errors.
Finally, the Weather-dynamics-aware GCN constructs a dynamic adjacency matrix based on deep features extracted by the MSA-KAN layer. This allows for collaborative inference in visibility-restricted areas, where foggy conditions make it difficult to accurately detect people.
Experiments conducted on four public datasets demonstrate the effectiveness of this new approach. In dense fog scenarios, the proposed method achieves a 12.2-27.5% reduction in mean absolute error (MAE) compared to mainstream algorithms. Furthermore, the framework exhibits robustness under varying fog concentrations and weather conditions.
The integration of physical principles with deep learning techniques offers a promising solution for crowd counting in challenging environments. This approach has significant implications for applications such as intelligent monitoring, urban management, and emergency response systems. By accurately detecting and counting people in foggy conditions, this technology can improve public safety and enhance decision-making processes in various fields.
Cite this article: “Fog-Friendly Crowd Counting: Combining Physics and Deep Learning for Improved Accuracy”, The Science Archive, 2025.
Fog, Crowd Counting, Deep Learning, Atmospheric Scattering, Physics-Guided, Multi-Scale Attention, Graph Convolutional Network, Visibility-Restricted Areas, Public Safety, Emergency Response Systems.
