PHGDiff: physics-enhanced hypergraph diffusion model for crowd trajectory prediction

Accurate pedestrian trajectory prediction is crucial for the safety of autonomous driving, intelligent transportation, and human-robot collaboration systems. However, the task remains challenging due to the high uncertainty of human behavior and complex social interactions. Existing models struggle to balance trajectory diversity and physical plausibility.This paper proposes a novel physics-enhanced hypergraph diffusion model (PHGDiff) to address this challenge. PHGDiff integrates three key innovations: a dynamic hypergraph mechanism that captures high-order group interactions via hyperedges, improving the model's ability to represent social behaviors such as cooperation and avoidance; a Transformer-based conditional diffusion framework that combines the strengths of diffusion models in multimodal generation and the temporal dependency modeling of Transformers, enabling stable and diverse trajectory synthesis; a noise-aware physical constraint scheduler that incorporates both social force-based dynamic soft constraints and kinematic hard constraints. This scheduler adjusts constraint weights based on noise levels, preserving diversity in early training and enhancing physical consistency in later stages.Extensive experiments on ETH and UCY datasets demonstrate the effectiveness and robustness of PHGDiff, achieving superior performance in trajectory prediction accuracy, diversity, and physical plausibility. This work provides a promising solution for physically consistent and socially aware multimodal trajectory prediction in complex real-world environments.