Learning to predict trajectories with destinations from massive vessel data

Accurate and real-time ship trajectory prediction is a premise for high-stake tasks such as risk reduction, route planning, energy saving, etc., and becomes more feasible based on the processing of AIS data with sophisticated algorithms, so as to ensure high-standard navigation by providing efficient trajectory-based maritime traffic management. In contrast to current prevailing research striving to improve short-term prediction accuracy, this paper focuses on whereabouts estimation in order to improve longer-term predictions for vessels. Taking the meaningful whereabouts as implicit destinations, the novel Destination-Guided Trajectory Prediction (DGTP) model is proposed, which employs a cascaded Seq2Seq architecture with BiGRU to simultaneously predict both vessel destination and trajectory. Trajectory Alignment Loss (TAL) is also introduced to encourage precise matching between the predicted and true trajectories in optimizing the DGTP model. Experiments conducted on a large volume of AIS data demonstrate that both destination prediction and TAL loss can independently improve trajectory prediction performances. Moreover, the synergistic combination of destination prediction and TAL within the DGTP model leads to substantial accuracy enhancements, demonstrating the promising results in long-term prediction.