SWIFT: A Small-World Interaction Framework for Flow-Aware Trajectory Prediction in Autonomous Driving

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Accurate trajectory prediction is crucial for autonomous driving safety.
• Current data-driven models struggle with generalization due to lack of structural priors.
• Modeling dynamic and context-dependent interactions among traffic agents is challenging.

Purpose of the Study:

• To propose SWIFT (Small-World Interaction Framework for Trajectory prediction), a unified framework for trajectory prediction.
• To integrate small-world network theory and traffic flow theory for improved interaction modeling.
• To enhance generalization and robustness of trajectory prediction models.

Main Methods:

• Introduced a Small-World Interaction Network for capturing local and global dependencies.
• Developed a Flow Regime Encoder to adapt interaction structures to traffic states.
• Utilized a multi-relational graph module for explicit encoding of agent relationships.

Main Results:

• SWIFT consistently outperformed strong baselines in prediction accuracy on nuScenes, MoCAD, and NGSIM datasets.
• Demonstrated improved generalization to unseen locations and traffic regimes.
• Showcased robustness under noisy observations and strong performance with limited data.

Conclusions:

• The structure-aware design of SWIFT is effective for trajectory prediction in autonomous driving.
• Integrating network structure and traffic flow theory enhances model generalization and robustness.
• SWIFT offers a promising direction for developing more reliable autonomous driving systems.