Unified Multimodal Vessel Trajectory Prediction with Explainable Navigation Intention

TL;DR

DI-MTP tackles the problem of reliable, explainable short-term vessel trajectory prediction in diverse encounter scenarios by decomposing multimodality into sustained and transient intentions. The method combines a sustained-intention tree, built from historical data, with a global transient intention optimization powered by a destination-aware CVAE and non-local attention to ensure coherent, scenario-wide predictions. It achieves significant ADE and FDE improvements over both deterministic and multimodal baselines on real AIS datasets, and offers explicit modal explanations via attention-derived prototypes. The approach enhances safety and interpretability in maritime navigation with practical potential for real-time integration in intelligent maritime systems, while leaving room for future improvements in interpretability and regulatory alignment.

Abstract

Vessel trajectory prediction is fundamental to intelligent maritime systems. Within this domain, short-term prediction of rapid behavioral changes in complex maritime environments has established multimodal trajectory prediction (MTP) as a promising research area. However, existing vessel MTP methods suffer from limited scenario applicability and insufficient explainability. To address these challenges, we propose a unified MTP framework incorporating explainable navigation intentions, which we classify into sustained and transient categories. Our method constructs sustained intention trees from historical trajectories and models dynamic transient intentions using a Conditional Variational Autoencoder (CVAE), while using a non-local attention mechanism to maintain global scenario consistency. Experiments on real Automatic Identification System (AIS) datasets demonstrates our method's broad applicability across diverse scenarios, achieving significant improvements in both ADE and FDE. Furthermore, our method improves explainability by explicitly revealing the navigational intentions underlying each predicted trajectory.

Figures (7)

• Figure 1: Comparison of vessel trajectory prediction methods, where orange dashed lines indicate observed trajectories, black solid lines show actual trajectories, and blue/red dashed lines represent predicted trajectories for two vessels. (a) Deterministic prediction generates a single fixed trajectory, emphasizing route selection while disregarding vessel encounters. (b) Current multimodal prediction, shown with a light blue area indicating trajectory variation range, achieves predictions with equal probabilities and limited modal diversity. (c) Our proposed multimodal intention-aware prediction offers diverse, realistic trajectories with differentiated probabilities (indicated by line thickness) and explicit navigation intentions (represented by icons).
• Figure 2: Overall of DI-MTP. Sustained Intentions are extracted from historical trajectories and matched with observed trajectories to construct sustained intention trees. Next, the relationships between sustained and transient intentions are captured by modeling trajectory destinations. The overall encoding passes through a non-local attention mechanism to further strengthen the relationships between vessels in the global context, and finally, the decoding obtains the final prediction.
• Figure 3: Trend curves of ADE and FDE under varying sequence settings: (a) observation lengths from 2 to 10 timesteps (prediction length fixed at 10); (b) prediction lengths from 8 to 14 timesteps (observation length fixed at 6). Lower values indicate better performance.
• Figure 4: Trend curves of ADE/FDE for varying numbers of sustained intentions $k$ in (a) and varying sampling numbers $n$ in (b). Lower metric values are better.
• Figure 5: Sustained intentions in diverse scenarios, where yellow lines indicate observed trajectories, black lines represent ground truth, and red, green, and blue dashed lines show the top 3 sustained intentions for each vessel with their MMSI in legend, respectively. (a) Non-encounter scenario. (b) Two vessels encounter scenario. (c) Three vessels encounter scenario.