Hazy Pedestrian Trajectory Prediction via Physical Priors and Graph-Mamba

Jian Chen, Zhuoran Zheng, Han Hu, Guijuan Zhang, Dianjie Lu, Liang Li, Chen LyuarXiv 2025 · arXiv · OpenReview

Introduction

Predicting pedestrian trajectories is a fundamental requirement for autonomous driving and smart city safety systems. However, adverse weather conditions like heavy haze present a significant challenge, causing visual degradation and altering typical human social behaviors. Standard trajectory prediction models often fail in these environments because they lack physical grounding and struggle with the computational complexity of modeling dense social interactions. To address these issues, researchers have developed FAPTP (Physical-prior-based Atmospheric and Pedestrian Trajectory Prediction), a framework that combines atmospheric physics with state-of-the-art sequence modeling.

The FAPTP Framework

The FAPTP architecture is designed to mitigate the effects of haze by integrating physical priors with an efficient "Graph-Mamba" architecture. The system is built upon three primary pillars that handle feature extraction, temporal sequence modeling, and social interaction.

The FAPTP architecture consists of the PhyFusion module for physical parameter estimation, MambaST-Sequences for temporal modeling, and DynaHetero-Net for social interaction modeling.

1. PhyFusion Module

Instead of relying on standard image dehazing, the PhyFusion module utilizes the Koschmieder atmospheric scattering model to estimate physical parameters such as scene depth, scattering coefficients, and global illumination. By decoupling haze concentration from light degradation, the module extracts "haze-invariant" features that remain stable even as visibility decreases.

2. MambaST-Sequences

To capture the temporal evolution of pedestrian movement, FAPTP employs a variant of the Mamba State Space Model. The researchers developed an adaptive scanning mechanism that is 78% faster than the native Mamba architecture. This allows the model to capture long-range dependencies with linear complexity, ensuring high-speed inference suitable for real-time applications.

3. DynaHetero-Net

Social interaction patterns change in low visibility; humans typically reduce their social perception range. The DynaHetero-Net is a heterogeneous graph attention network that models individuals and groups as distinct nodes. It uses a "hazy perceptual modulation function" to dynamically adjust the influence of social interactions based on current visibility levels and pedestrian proximity.

Experimental Results

The model was rigorously tested on the ETH/UCY dataset, which was augmented with synthetic haze at four intensity levels, ranging from clear conditions ($\beta=0$) to heavy haze ($\beta=2.0$, visibility ~30m).

Quantitative comparison of FAPTP against state-of-the-art models across different datasets and haze intensities (beta). FAPTP consistently achieves the lowest minADE/minFDE (bolded).

Accuracy and Robustness

In dense haze scenarios, FAPTP demonstrated significant improvements over existing state-of-the-art (SOTA) models. It reduced the minADE (Average Displacement Error) by 37.2% and the minFDE (Final Displacement Error) by 41.5%. While competing models like TimeMachine saw error rates spike by over 50% when moving from moderate to heavy haze, FAPTP maintained superior stability, proving its robustness in extreme weather.

Real-time Efficiency

FAPTP is optimized for deployment on edge devices, achieving a throughput of 494.5 FPS. This high frame rate, combined with its predictive accuracy, makes it a viable solution for the real-time safety requirements of autonomous vehicles operating in adverse environments.

Conclusion

FAPTP introduces a new paradigm for pedestrian trajectory prediction by grounding deep learning in physical reality. By combining atmospheric scattering priors with the efficient Graph-Mamba architecture, it provides a reliable and computationally efficient solution for intelligent transportation systems. Future research may explore extending this framework to other weather conditions, such as rain or snow, and incorporating multi-modal sensor fusion to further enhance perception in zero-visibility scenarios.

Summary updated

Abstract

To address the issues of physical information degradation and ineffective pedestrian interaction modeling in pedestrian trajectory prediction under hazy weather conditions, we propose a deep learning model that combines physical priors of atmospheric scattering with topological modeling of pedestrian relationships. Specifically, we first construct a differentiable atmospheric scattering model that decouples haze concentration from light degradation through a network with physical parameter estimation, enabling the learning of haze-mitigated feature representations. Second, we design an adaptive scanning state space model for feature extraction. Our adaptive Mamba variant achieves a 78% inference speed increase over native Mamba while preserving long-range dependency modeling. Finally, to efficiently model pedestrian relationships, we develop a heterogeneous graph attention network, using graph matrices to model multi-granularity interactions between pedestrians and groups, combined with a spatio-temporal fusion module to capture the collaborative evolution patterns of pedestrian movements. Furthermore, we constructed a new pedestrian trajectory prediction dataset based on ETH/UCY to evaluate the effectiveness of the proposed method. Experiments show that our method reduces the minADE / minFDE metrics by 37.2% and 41.5%, respectively, compared to the SOTA models in dense haze scenarios (visibility < 30m), providing a new modeling paradigm for reliable perception in intelligent transportation systems in adverse environments.
BibTeX
@misc{chen2025hazy,
  author = {Jian Chen and Zhuoran Zheng and Han Hu and Guijuan Zhang and Dianjie Lu and Liang Li and Chen Lyu},
  title = {{Hazy Pedestrian Trajectory Prediction via Physical Priors and Graph-Mamba}},
  year = {2025},
  eprint = {2509.24020},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.24020},
}
Figure