A Learning-Based Pedestrian Flow Prediction Approach with Diffusion Behavior

Transportation Research Part C: Emerging Technologies 2025

Authors: Weiming Mai, Dorine Duives, Panchamy Krishnakumari, Serge Hoogendoorn

Diffusion-based pedestrian flow prediction framework — Figure: Overview of the diffusion-based pedestrian flow prediction framework.

Abstract

Accurately predicting pedestrian flows is essential for effective crowd management in public spaces such as train stations, airports, and event venues. Existing data-driven methods typically treat links in a pedestrian network independently or rely solely on spatial adjacency, often neglecting the underlying physical process of how crowds propagate through the network. In this paper, we propose a learning-based approach that explicitly models crowd diffusion behavior - the process by which pedestrian flows spread from origin to destination through intermediate links over time. Our method integrates a diffusion-inspired graph representation with a Graph Neural Network (GNN) architecture, capturing both the spatial connectivity and temporal propagation patterns of pedestrian flows. Furthermore, we introduce an online learning mechanism that allows the model to adapt to changing demand patterns in real time. Experiments on both synthetic and real-world datasets demonstrate that our approach outperforms state-of-the-art baselines.

Key Contributions

1 A novel diffusion-inspired graph representation that captures how pedestrian flows propagate through network links over time.
2 Integration with Graph Neural Networks (GNNs) for spatiotemporal forecasting that respects the physical diffusion process.
3 An online learning mechanism for real-time adaptation to changing demand patterns.
4 Extensive experiments on synthetic and real-world datasets showing superior performance.

Method Overview

Diffusion Graph

Models pedestrian flow propagation as a diffusion process on a network graph, where flows spread from origins to destinations through intermediate links with time delays.

GNN + Online Learning

A GNN-based architecture learns spatiotemporal patterns from the diffusion graph, with online learning enabling continuous adaptation to real-time observations.

Citation

@article{mai2025diffusion,
  title={A Learning-Based Pedestrian Flow Prediction Approach
         with Diffusion Behavior},
  author={Mai, Weiming and Duives, Dorine and Krishnakumari,
          Panchamy and Hoogendoorn, Serge},
  journal={Transportation Research Part C},
  year={2025},
  publisher={Elsevier}
}