Parallel Transformer-GAT with Intention Scoring for Pedestrian Trajectory Prediction

Abstract

The inherent uncertainty and dynamic interactivity of pedestrian motion in complex, crowded scenarios pose fundamental challenges for the safe decision-making of autonomous driving and intelligent social robotics. Specifically, pedestrian trajectory prediction must accurately integrate individual subjective motion intentions with objective environmental constraints. However, current methods often struggle to effectively align highly uncertain subjective motion intentions with dynamic objective constraints. To address this, we propose a parallel Transformer-GAT prediction architecture with a route-specific feature scoring mechanism. To effectively navigate high uncertainty and complex interactivity, a framework with dual Transformer and GAT modules is developed to extract individual subjective motion intentions and dynamically capture spatial interaction constraints in parallel. Building upon these features, an endpoint-aware Farthest Point Sampling initialization strategy is applied to ensure the diversity of candidate modes. Ultimately, a route-specific feature scoring mechanism is designed that achieves precise confidence assessment for multi-modal candidate trajectories through spatial geometric alignment in physical directions. Extensive experiments on the ETH and UCY datasets demonstrate that the proposed method achieves an average ADE of 0.24 and an FDE of 0.41, outperforming most state-of-the-art pedestrian trajectory prediction methods. Ablation studies further validate the effectiveness of each component.