Goal-driven long-term marine vessel trajectory prediction with a memory-enhanced network

Abstract

Enhancing the precision of marine vessel trajectory prediction (VTP) is crucial for collision avoidance, intelligent navigation, and crisis alert in maritime safety. Most RNN-based methods typically face memory weakening issues during long-sequence propagation, leading to the discarding of some key features and significant predictive error accumulation over extended time intervals. Moreover, they struggle to forecast those complex trajectories involving abnormal maneuvers such as sudden acceleration or deceleration, sharp turns, or U-turns, resulting in poor generalization capabilities. To address these pivotal challenges, this paper proposes a novel Memory-Enhanced Network (MENet) for VTP, catering to intricate sailing intention modeling with long-term motion pattern perception. Specifically, we design an embeddable memory-enhanced block (MEB) that adaptively aggregates memory vectors across multiple temporal scales to assist in better prediction without disrupting the original backbone structure. Also, a goal-driven vessel trajectory decoder (GD-VTD) is developed to facilitate reliable model inferences by combining vessel type and destination variables as guidance information. Furthermore, we reconstruct the traditional loss function based on relative distance metrics, incorporating predicted headings into the optimization process to generate consistent trajectories that comply with realistic vessel dynamics. Ultimately, MENet could learn diverse sailing intentions by assembling the above parts to predict long-term marine vessel trajectories. Extensive experimental results on Automatic Identification System (AIS) datasets from three coastal regions in the US demonstrate that our model exhibits superior accuracy and robustness compared to other baselines. Specifically, on the Everglades Port (EP) dataset, our method reduces MAE, RMSE, and MAPE errors by 7.25%, 7.82%, and 7.62%, respectively, compared to the existing best results during this experiment. This is another piece of evidence for the effectiveness of goal-driven trajectory prediction in real-world maritime settings.