Foggy Crowd Counting: Combining Physical Priors and KAN-Graph

Yuhao Wang, Zhuoran Zheng, Han Hu, Dianjie Lu, Guijuan Zhang, Chen LyuarXiv 2025 · DOI · OpenReview

Overview

Crowd counting is a critical tool for urban management and public safety. However, environmental factors like heavy fog significantly degrade the performance of standard computer vision models. This paper introduces the Foggy Crowd Counting Network (FCCN), a novel framework that blends the laws of physics with cutting-edge neural architectures to maintain high accuracy in adverse weather. FCCN achieves a 12.2% to 27.5% reduction in Mean Absolute Error (MAE) compared to current state-of-the-art methods in dense fog scenarios.

The Challenge of Foggy Environments

Fog introduces three major hurdles for traditional crowd-counting systems:

  1. Non-uniform Attenuation: Fog does not blur an image evenly; it depends on the depth of field, meaning distant objects are more obscured than near ones.
  2. Feature Degradation: Standard convolutional neural networks (CNNs) struggle to extract meaningful patterns from low-contrast, blurry regions.
  3. Loss of Spatial Context: When fog hides the "connective tissue" of a scene, models lose the ability to understand how different parts of a crowd relate to one another.

Existing models often treat fog as simple noise, but FCCN argues for modeling the physical process of light scattering to effectively "see" through the haze.

Methodology

The FCCN framework consists of three synergistic modules designed to handle atmospheric degradation:

The FCCN architecture integrates physics-based dehazing with advanced neural structures like MSA-KAN and GCNs.
  1. Physics-Guided Module: Instead of simple pre-processing, this module uses a differentiable version of Koschmieder’s atmospheric scattering theory. It estimates transmittance and atmospheric light to restore a "dehazed" version of the scene, using a U-Net for refinement and guided filtering to preserve sharp edges.
  2. Multi-Scale Adaptive Kolmogorov-Arnold Network (MSA-KAN): Departing from standard CNNs, KANs use learnable activation functions (B-splines) on their edges rather than fixed functions on nodes. This three-layer progressive KAN is significantly better at fitting the complex, non-linear patterns found in degraded, foggy features.
  3. Weather-Aware Graph Convolutional Network (GCN): To handle spatial relationships, the model treats image regions as nodes in a graph. It dynamically builds an adjacency matrix that adjusts based on fog density, allowing the network to transfer information from high-visibility regions to low-visibility ones.

Experimental Setup and Datasets

The authors evaluated FCCN on four major public crowd-counting datasets: ShanghaiTech A, UCF-QNRF, JHU-Crowd++, and NWPU-CROWD. Because these datasets are typically clear, the authors constructed a hierarchical haze benchmark by systematically synthesizing four levels of fog (from "no haze" to "severe haze") using physical scattering equations.

Key Results and Insights

FCCN consistently outperformed eight other state-of-the-art methods (including SFCN, BL, and MAN) across all datasets and fog levels.

Visual comparison of density maps and crowd counts across different models in clear and foggy conditions.
  • Accuracy and Robustness: In moderate haze, FCCN achieved an MAE of 85.8, significantly lower than the next best model, AWCC-Net (96.3). While traditional models saw error rates surge by over 35% when fog density doubled, FCCN’s error only increased by 22%.
  • KANs over CNNs: KAN layers reduced feature confusion in dense fog by 31% compared to traditional convolutional layers. Their ability to learn flexible mappings makes them uniquely suited for reconstructing details in blurry regions.
  • Joint Optimization: By integrating physics into the training loop rather than using it as a pre-processor, FCCN prevents the dehazing process from accidentally smoothing out tiny, distant human heads that are crucial for an accurate count.

Conclusion and Future Work

FCCN demonstrates that combining physical priors with flexible neural architectures like KANs and GCNs provides a robust solution for crowd counting in challenging weather. While the results are impressive, future work may focus on adapting the model to real-world fog variations (like smog or steam) and optimizing the computational complexity of KANs for real-time deployment. The authors suggest this approach could be extended to other adverse conditions, such as heavy rain or nighttime glare.

Summary updated

Abstract

Aiming at the key challenges of crowd counting in foggy environments, such as long-range target blurring, local feature degradation, and image contrast attenuation, this paper proposes a crowd-counting method with a physical a priori of atmospheric scattering, which improves crowd counting accuracy under complex meteorological conditions through the synergistic optimization of the physical mechanism and data-driven.Specifically, first, the method introduces a differentiable atmospheric scattering model and employs transmittance dynamic estimation and scattering parameter adaptive calibration techniques to accurately quantify the nonlinear attenuation laws of haze on targets with different depths of field.Secondly, the MSA-KAN was designed based on the Kolmogorov-Arnold Representation Theorem to construct a learnable edge activation function. By integrating a multi-layer progressive architecture with adaptive skip connections, it significantly enhances the model's nonlinear representation capability in feature-degraded regions, effectively suppressing feature confusion under fog interference.Finally, we further propose a weather-aware GCN that dynamically constructs spatial adjacency matrices using deep features extracted by MSA-KAN. Experiments on four public datasets demonstrate that our method achieves a 12.2\%-27.5\% reduction in MAE metrics compared to mainstream algorithms in dense fog scenarios.
BibTeX
@misc{wang2025foggy,
  author = {Yuhao Wang and Zhuoran Zheng and Han Hu and Dianjie Lu and Guijuan Zhang and Chen Lyu},
  title = {{Foggy Crowd Counting: Combining Physical Priors and KAN-Graph}},
  year = {2025},
  eprint = {2509.24545},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.24545},
}
Figure