Ha-Yun Kim, "An Improved Method for Tracking of Maritime Obstacles Using Sensor Data", M.Sc. Thesis, Seoul National University, 2024.02.26
M.Sc. Thesis
2024.06.25 17:54
Ha-Yun Kim, "An Improved Method for Tracking of Maritime Obstacles Using Sensor Data", M.Sc. Thesis, Seoul National University, 2024.02.26
조회 수 930
| 첨부 '1' |
|---|
| Abstract | For the safe operation of ships, it is important to recognize and track nearby maritime obstacles accurately. Various types of sensors, such as RADAR (Radio Detection and Ranging), AIS (Automatic Identification System), cameras, etc., are used to detect maritime obstacles. However, getting accurate information about maritime obstacles can be difficult when sensor noise or data is missing. To compensate for this, we utilize a tracking algorithm based on the sensor data to track the status of the obstacle, such as trajectory, COG (Course Over Ground), and SOG (Speed Over Ground). There has been a lot of research on obstacle tracking methods, one of which is the EKF-based tracking method, which can track the state of an obstacle from measurements with noise. When computing the EKF (Extended Kalman Filter), the parameters of EKF may need to be optimized because the user usually sets parameters empirically. In addition, it requires an accurate system model of the target obstacle and a convergence for the tracking values to stabilize at the beginning of the tracking. There has been a lot of research on deep learning in various fields recently, and there is a lot of research on applying deep learning to obstacle tracking. Learning-based tracking methods using deep learning do not require a system model of the tracked object and track obstacles without initial convergence of tracking values. However, it requires a separate deep learning model training process and a large amount of data to train the model, and the accuracy is lower than the EKF-based tracking method. In this study, we improved the EKF-based tracking method by optimizing its parameters and implemented a learning-based tracking method using DNN (Deep Neural Network). In addition, we proposed a hybrid tracking method combining the EKF-based and learning-based tracking methods to compensate for the shortcomings of each method. The three tracking methods utilized in this study were verified using data obtained through field tests. The verification results showed that the learning-based tracking method reduced the SOG tracking accuracy by 11.47% compared to the EKF-based tracking method. Compared with the EKF-based tracking method, the hybrid tracking method has improved the convergence speed, and the tracking accuracy is reduced by 22.42% for COG and 42.05% for SOG. Therefore, these results demonstrate the hybrid tracking method can effectively track by compensating for the shortcomings of the other methods. |
|---|---|
| Publication Date | 2024-02-26 |
