New Transformer Boosts Vessel Trajectory Prediction Accuracy

A novel deep learning framework, the Continuous Regression Hybrid Transformer (CRHT), has been proposed to enhance the accuracy of vessel trajectory prediction. This model is specifically designed to overcome common challenges in maritime forecasting, such as geographic data bias and the need for realistic navigational predictions. CRHT integrates 1D convolutional layers to extract local kinematic features from Automatic Identification System (AIS) data, alongside a multi-head attention mechanism to capture global temporal context. To mitigate the issue of spatial data imbalance, which often leads to underrepresentation of rare maneuvers, the framework incorporates an online K-means cluster sampling strategy during training. This ensures that the model is exposed to a diverse range of vessel behaviors. Empirical results demonstrate that CRHT achieves superior performance in short-term forecasting, particularly at the one-hour horizon, offering an optimal balance between prediction precision and the ability to track complex vessel maneuvers for real-time maritime surveillance.