STGDPM:Vessel Trajectory Prediction with Spatio-Temporal Graph Diffusion Probabilistic Model

TL;DR

The paper tackles vessel trajectory prediction under uncertainty and interactive multi-modality by introducing STGDPM, a framework that fuses Spatio-Temporal Graphs with Denoising Diffusion Probabilistic Models. It presents Traj-UGnet as the diffusion denoiser conditioned on historical trajectories and dynamic interaction graphs, enabling probabilistic, multimodal forecasting. Through extensive AIS data experiments, STGDPM achieves superior accuracy over strong baselines and demonstrates the ability to capture diverse future behaviors, including COLREG-compliant maneuvers. The approach promises practical benefits for autonomous navigation, collision avoidance, and maritime traffic safety in complex, dynamic environments.

Abstract

Vessel trajectory prediction is a critical component for ensuring maritime traffic safety and avoiding collisions. Due to the inherent uncertainty in vessel behavior, trajectory prediction systems must adopt a multimodal approach to accurately model potential future motion states. However, existing vessel trajectory prediction methods lack the ability to comprehensively model behavioral multi-modality. To better capture multimodal behavior in interactive scenarios, we propose modeling interactions as dynamic graphs, replacing traditional aggregation-based techniques that rely on vessel states. By leveraging the natural multimodal capabilities of diffusion models, we frame the trajectory prediction task as an inverse process of motion uncertainty diffusion, wherein uncertainties across potential navigational areas are progressively eliminated until the desired trajectories is produced. In summary, we pioneer the integration of Spatio-Temporal Graph (STG) with diffusion models in ship trajectory prediction. Extensive experiments on real Automatic Identification System (AIS) data validate the superiority of our approach.

Figures (6)

• Figure 1: STGDMP framework.
• Figure 2: Architecture of our Traj-UGnet module. The color of the arrow corresponds to the color of the function module.
• Figure 3: From left to right are Caofeidian Waters and Tianjin Port. Tianjin, with its moderate depth and urban location, handles medium-sized container ships, tankers, and ro-ro vessels, making it a key multimodal trade hub for northern China. In contrast, Caofeidian’s deep-water facilities support vessels over 200,000 DWT, focusing on bulk commodities like coal, iron ore, and crude oil, with infrastructure for large-scale handling and transshipment.
• Figure 4: We compared the prediction results of various methods in complex scenarios. The left image shows a turning and acceleration scenario, while the middle and right images depict turning scenarios.
• Figure 5: Behavioral multi-modality of trajectories predicted by our model.