Abstract
Accurate tracking of the position and motion states of maritime obstacles is essential for the safe autonomous operation of ships. However, sensor measurements obtained from RADAR (RAdio Detection And Ranging) and cameras often contain uncertainties, such as noise, missing observations, and outliers, which make stable maritime obstacle tracking difficult. Although conventional EKF (Extended Kalman Filter)-based tracking methods, including IMM-EKF (Interacting Multiple Model-Extended Kalman Filter), have been widely used to address this problem, they require parameter tuning when maritime environmental conditions change and remain dependent on predefined motion models. To overcome these limitations, this study proposed a method of learning-based tracking for maritime obstacles using a Transformer. First, a synthetic training dataset was generated by incorporating the distance-dependent uncertainty characteristics of real-world sensor measurements. Then, a Transformer-based tracking model architecture was designed to learn the time-series characteristics of sensor measurements and to track the position, SOG (Speed Over Ground), and COG (Course Over Ground) of maritime obstacles. The proposed method was evaluated using MAE (Mean Absolute Error) and compared with the optimized IMM-EKF, whose parameters were optimized under the maritime environmental conditions of Scenario A. In Scenario A, the proposed method achieved MAEs of 6.66 m, 1.76 knots, and 10.70° for position, SOG, and COG, respectively, whereas the optimized IMM-EKF yielded MAEs of 6.73 m, 1.35 knots, and 14.89°, respectively. These results indicate that the proposed method achieved comparable position tracking performance and a lower MAE for COG, whereas the optimized IMM-EKF achieved a lower MAE for SOG. In Scenario B, under the changed maritime environmental conditions, the proposed method achieved MAEs of 11.07 m, 1.65 knots, and 16.52° for position, SOG, and COG, respectively, reducing the MAEs by approximately 31.79%, 16.67%, and 32.71% compared with the optimized IMM-EKF. These results indicate that the proposed method maintained robust tracking performance, as reflected by lower average MAEs, under the changed maritime environmental conditions without additional parameter tuning or redefinition of the motion model.
