SocialVessel: VAE-inspired Multi-Vessel Gated Recurrent Units for Vessel Trajectory Prediction in Port Areas

Accurate and robust trajectory prediction models are crucial for ensuring the safe navigation of autonomous vessels. To meet this demand, we develop SocialVessel, a novel framework that leverages a VAE-inspired archtitecture with Gated Recurrent Units to enhance vessel trajectory prediction in port areas. By considering the target vessel and its surroundings, SocialVessel simultaneously infers multiple predicted trajectories and employs clustering to determine the most probable destination and trajectory. Through the incorporation of an attention mechanism, the model effectively captures the influence of surrounding vessels by encoding essential features such as distance, DCPA, and TCPA. Additionally, a VAE-inspired framework aligns the predictive distribution with the observed distribution during training. Experimental results demonstrate the superior performance of SocialVessel compared to existing models, showcasing its potential to enhance the navigation capabilities of autonomous vessels.