Enhancing risk perception by integrating ship interactions in multi-ship encounters: A Graph-based Learning method

The navigation safety of autonomous surface ships depends on risk perception and avoidance in advance, which is based on accurate trajectory prediction of other ships. Sequential neural networks in deep learning have demonstrated reliable predictions in navigation scenarios with limited multi-ship interactions. However, accurately predicting trajectory changes caused by ship interactions remains challenging, as these predictions are based on mutually independent historical trajectories. In multi-ship encounters, trajectory predictions that lack interaction considerations can cause subsequent risk perception away from the actual future risk, thereby compromising navigation safety. In this study, we propose a method, the Graph-based Learning model for Risk Perception (GLRP), for risk perception based on interactive trajectory prediction. It introduces a variational graph auto-encoder to simulate the uncertain actions of ships in interactive environments, and takes the self-attention block to learn global time dependencies. GLRP establishes a learning channel from ship interactions to ship trajectories, allowing predictions based on exchanged trajectory inputs. The experiments indicate that GLRP reduces the distance to the closest point of approach error by 5. 45% and the time to the closest point of approach error by 4. 85% compared to individual sequence models. It improves navigation safety by enhancing the reliability of risk perception. The implementation code of this work is available at: https://github.com/KaysenWB/RESS_GLRP.