Yeongmin Jo, Myung-Il Roh, Hye-Won Lee, Donghun Yu, "A Ship Tracking Method under Dynamic Characteristic Changes with LSTM", Proceedings of G-NAOE 2022, Changwon, Korea, 2022.11.06-10
|For the safe operation of a ship, it is necessary to detect and track nearby ships, called target ships, accurately. To track the target ships, Kalman filters (KF) have been widely used. However, when the system model of the KF, which predicts the motion of the system, may be different from the actual motion of the target ships, the accuracy could decrease. To solve this problem, the Interactive Multi-Model (IMM) method can be adopted, which follows several system models at the same time, but the number of system models that the IMM implements is limited. Therefore, a method that can analyze the actual dynamic behavior of the target ships and track accurately the target ships based on historical data is required. Recently, deep learning has been applied to tracking methods to solve existing problems and improve accuracy. For tracking, the Long-Short Term Memory (LSTM) method is being mainly applied and shows better performance in the marine environment where the motion characteristics of the target ships could change.
In this study, a tracking method based on deep learning is proposed to track nearby target ships. First, we construct a tracking filter using the LSTM-KF method, which combines LSTM and traditional KF. In the LSTM-KF, the system model and system noise of the KF are trained by previous data and predicted using deep learning. There is also a method of constructing an LSTM model that directly produces the track of the target ships as output data. After constructing the tracking filter using LSTM, the tracking filter is trained using the ship’s navigation data. In this study, the proposed method is compared with the tracking results using the traditional KF, and it was confirmed that it works effectively.