Abstract
Maritime surveillance radar technologies are widely employed for detecting and tracking vessels within the maritime domain; however, they exhibit limitations in accurately classifying and identifying ships. Such inaccuracies can lead to significant challenges, including compromised safety and security, inefficient resource management, and reduced situational awareness. This thesis addresses these limitations by developing and evaluating a ship identification system that integrates deep multimodal neural networks, the Extended Kalman Filter (EKF), and blockchain technology. One key issue arises when a coastal surveillance radar detects a ship and assigns it a target identifier while another distant coastal radar or an Automatic Identification System (AIS) station detects the same ship and assigns a different target identifier.
The developed system connects radar and AIS trajectories, enabling the identification of a ship even when it is first picked up by a radar and later detected by an AIS station at a different location. The system models the trajectory using the first radar's data to predict the future position of the ship, comparing this predicted trajectory with those recorded by the second radar or AIS station. If a match is found, the ship is confirmed as the same one detected initially. Future trajectories are predicted using Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and the EKF, with trajectory fusion carried out by another EKF. Each predicted and fused trajectory is compared to find the best match, with the closest trajectory selected as the most accurate. Blockchain technology is applied to ensure the integrity of the data collected using radar and AIS systems.
The results demonstrate that each model performs differently depending on the trajectory it aims to match. The discrete Fréchet distance (DFD) was used as the primary matching technique, with other methods such as Dynamic Time Warping (DTW) and Siamese networks reviewed for their applicability. The system was tested by comparing radar and Automatic Identification System (AIS) trajectories and further evaluated using deep neural networks and deep reinforcement learning algorithms, such as Deep Q-Networks (DQN), to assess their fit to the problem. Radar and AIS data collected within the South African Exclusive Economic Zone (EEZ) in August 2022 formed the basis for these tests...