SIMULATION-BASED DATA COLLECTION AND DEEP LEARNING FOR DISTRESS SHIP DETECTION USING DRONES FOR MARITIME SEARCH AND RESCUE

Drones and deep learning are increasingly being used to quickly detect distress at sea, but the lack of data on distress ships limits detection. In this study, we develop a simulator for marine ship accidents to construct a dataset and train an object detection model to detect distress ships. When developing the simulator, we focused on sinking and capsizing ships. The drone was then used to construct a dataset of ships in distress. We selected the YOLOv8 object detection model for its accuracy and real-time performance, and we trained it using the constructed dataset, evaluating its performance based on various indicators. It was possible to identify and detect ships and distress ships, and all were detected without any missing ships in the test data. As a result, a high accuracy of mAP 0.969 was achieved. The results of this study show that simulation-based datasets can be useful for distress ship detection. In the future, if various marine environments and ships are implemented in the simulation to obtain learning data, it is expected to help minimize human casualties by quickly detecting distress ships in wide oceans.