A transformer framework for vessel trajectory prediction in complex maritime scenarios with an original vessel trajectory dataset

Abstract

In maritime intelligent collision avoidance, accurate multi-step ship trajectory prediction is essential for assessing collision risk and making proactive decisions. Conventional artificial intelligence methods typically rely on low-dimensional kinematic inputs, which often fail to capture complex maneuvering intentions and suffer from severe error accumulation over extended prediction horizons. To address these limitations, this paper proposes the Segmented Autoregressive Global Attention Mechanism Transformer (SAGM-TF) as an optimized trajectory prediction framework. Specifically, the framework integrates three key technical components: an optimized kinematic encoding module, a segmented autoregressive decoding strategy, and a cross-segment global attention mechanism. To capture vessel maneuvering characteristics more accurately, a dedicated trajectory dataset is constructed from self-collected navigation data, and the input feature space is expanded to include heading angle and angular velocity. Experimental results demonstrate that the proposed framework significantly mitigates cumulative errors and achieves stable, reliable multi-step trajectory prediction over a ninety-second time horizon in complex maritime environments. Compared with traditional methods, the proposed framework exhibits superior performance in trajectory fidelity and robustness, providing a reliable foundation for autonomous ship navigation systems based on artificial intelligence.