Vision Meets Constraints: Correcting Stray Trajectory by Utilizing Road Constraints

Current trajectory prediction models using Graph Neural Networks (GNNs) suffer from a critical failure pattern we term “Stray,” where predicted trajectories irrationally diverge into undrivable areas due to ignored map constraints. To resolve this, we propose a novel prediction framework that explicitly integrates road constraint information to eliminate physically unfeasible paths. Our approach introduces two key technical innovations: (1) a Constraint Feature mechanism that encodes drivable area boundaries to directly suppress Strays, and (2) a Prioritized Feature Pipeline that strategically processes Agent-Lane interactions distinctively from Agent-Agent interactions, predicting motion goals accurately. Experimental evaluations on the Argoverse Motion Forecasting dataset demonstrate the superiority of our method. Quantitatively, our approach achieved significant improvements in motion goal accuracy, securing the 4th rank on the leaderboard in terms of Miss Rate (as of January 8, 2026). Specifically, our model outperformed the strong baseline (HPNet) with a Miss Rate of approximately 0.10 and a minFDE of 0.97m in validation, proving that explicitly modeling road constraints and prioritizing lane features effectively mitigate Strays and contribute to safer decision making.