Trajectory prediction of seagoing ships in dynamic traffic scenes via a gated spatio-temporal graph aggregation network

Abstract

Accurate ship trajectory prediction is essential in maritime traffic control and safety, requiring the consideration of complex spatial and temporal dependencies within trajectory data. Most existing studies employ GNN and RNN to tackle this issue. However, (1) GNN-based approaches heavily rely on predefined graph structures to capture the static properties of ships, overlooking the dynamic interactions between them. (2) RNN-based strategies fail to capture long-term temporal dependencies. Also, the complex recursive architectures impede their ability to model local (short-term) temporal dependencies, leading to inefficient inference. To address the aforementioned challenges, this paper presents a novel Gated Spatio-Temporal Graph Aggregation Network (G-STGAN) for ship trajectory prediction, which comprises several key components. Firstly, a ship spatial gating encoder (SSGE) combines graph convolutional networks (GCN) and transformers to simulate dynamic and static spatial interactions, thereby improving prediction performance. Additionally, we design a ship temporal gating encoder (STGE), which utilizes a gated transformer (GT) and temporal convolution (TC) to capture both short- and long-term temporal dependencies. Ultimately, the spatial and temporal features obtained from the SSGE and STGE modules, respectively, are aggregated through a temporal convolutional network (TCN) to perform downstream trajectory prediction tasks. Experimental evaluations on three real-world Automatic Identification System (AIS) datasets demonstrate that G-STGAN achieves competitive prediction performance in terms of accuracy and robustness.