Deep Learning Based Underwater Object Detection and Hovering Control of Unmanned Underwater Vehicle

Continuous research had been conducted on the operating systems that can manufacture, maintain, and repair marine structures. When Divers are put into exploration of offshore structures, there are limitations such as high water pressure, diver's desease, and activity, so unmanned underwater vehicles(UUV) are used the underwater environment. However, UUVs also have a problem of drifting by ocean currents, waves, and wind. and it is difficult to know its location because GPS can’t be used under the water. Therefore, in this paper, the camera is used to find the real-time location and to perform hovering control of the UUV. This paper presents an object detection algorithm using YOLOv2, which has high real-time performance among deep learning-based object detection algorithms, and used it to detect specific types of objects in camera images in real time and to obtain location information. In order to identify objects in various underwater environments, it was trained with variety of conditions such as the illumination, distance, and presence or absence of obstructions by taking these measures, detection stability was enhanced. The distance between the camera and the object was measured using the relationship between the focal length of the camera and the shape and size of the object. The hovering control between the UUV and the object was implemented using the proposed method. For hovering control, a six degrees-of-freedom UUV was designed and constructed. For robust system, redundant thrusters were deployed, and the thruster arrangement matrix was studied. The control system for the posture control using the attitude heading reference system sensor and the depth sensor was set up, and it was implemented using the camera information. Sensor reliability was verified through individual sensor performance tests, and posture control experiments and hovering control experiments were conducted in sea. The ultra short base line sensor was used to verify the relative position estimation performance of the proposed system. Through the sea experiment, the attitude control confirmed the RMS error of roll, pitch, and yaw within 1° and the RMS error of depth within 1 cm. For hovering control, it was confirmed that hovering control using a camera was successful through the experimental results: RMS error of roll, pitch, and yaw within 2°, and the RMS distance error within 4 cm.1. 서 론 1.1 연구 배경 1 1.2 연구 목적 3 1.3 논문 구성 3 2. 무인잠수정 구성 2.1 구조 5 2.2 전원부 10 2.3 제어부 13 2.4 센서부 14 3. 객체 탐지 알고리즘 3.1 YOLOv2 19 3.2 레이어 구성 및 학습 옵션 25 3.3 수중물체 인식 학습 29 4. 제어 시스템 4.1 ROV 동역학 32 4.2 추진기 배치 및 추력 40 4.3 자세 제어 45 4.4 객체탐지 시스템을 이용한 위치제어 알고리즘 48 5. 센서 성능 실험 및 실해역 실험 5.1 센서 성능 실험 53 5.1.1 USBL 실험 53 5.1.2 AHRS 실험 53 5.1.3 Depth 실험 55 5.2 자세 제어 실험 57 5.3 무인 잠수정 호버링제어 실험 68 6. 결론 참고문헌 70 감사의 글 72Maste