IPLG: Intent-Propagated Lane Graph for Multimodal Trajectory Prediction at Urban Intersections

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Vehicle trajectory prediction at urban intersections is complex due to maneuver ambiguity, agent interactions, and intricate lane structures.
• Existing methods often face mode collapse or insufficient coverage of exit branches, impacting safety-critical applications.

Purpose of the Study:

• To propose the Intent-Propagated Lane Graph (IPLG) framework for multimodal vehicle trajectory prediction in UAV-observed urban intersections.
• To address semantic ambiguity at lane divergences and prevent predicted trajectories from deviating from valid lane regions.

Main Methods:

• Introduced a closed-loop interaction module for propagating agent intent cues within the lane graph topology.
• Developed a topology-aware diversified goal selection strategy to mitigate goal clustering across competing exit branches.
• Implemented a residual decoder to enforce local lane consistency during trajectory prediction.

Main Results:

• IPLG demonstrated consistent performance across multiple evaluation metrics on the XJROAD dataset.
• Achieved noticeable improvements in Miss Rate and Off-road Rate compared to existing methods.
• Validated generalizability on the nuPlan dataset, generating plausible trajectories across diverse intersection geometries.

Conclusions:

• The IPLG framework effectively enhances multimodal vehicle trajectory prediction at urban intersections.
• IPLG offers improved reliability and accuracy, particularly in safety-critical scenarios with complex driving behaviors.
• The proposed methods show strong generalizability across various road structures and driving conditions.