Interaction aware and multi-modal distribution for ship trajectory prediction with spatio-temporal crisscross hybrid network

Understanding the interactions between a ship and its surrounding ships enables effective trajectory prediction, which is critical to improving the safe navigation of autonomous ships. The prediction of future trajectories is a very challenging problem due to inherent uncertainties and complex spatiotemporal correlations between different ships. However, existing methods ignore the persistence and cross-domain nature of the influence between ships. To address the above challenges, an adaptive learning framework based on spatio-temporal crisscross hybrid network (STCNet) is proposed, which consists of two parts: spatio-temporal interaction aware and multi-modal trajectory prediction. Modeling temporal-dependent features, spatial interaction features and cross-domain features, and performs adaptive fusion to identify important features and capture all dynamic dependencies. Secondly, most methods only focus on the frequent modes of trajectories and cannot cover the actual paths of limited samples. Therefore, we design an augmented sampling method based on fusion knowledge and graph attention mechanism (KGS) to encourage exploration of trajectories in sparse areas of the sample space, and promote more accurate and reasonable future trajectory prediction. Experiments on the Ningbo-Zhoushan Port sea area dataset show that our method achieves better results than other methods.