부산대학교 도서관

This paper discussed cost-effective vision-based obstacle avoidance for Unmanned Aerial Vehicles (UAVs) operating in GPS-denied environments. The system combines the You Only Look Once (YOLO) architecture with stereo vision cameras (OAK-D Lite), a Raspberry Pi computer, and a flight controller unit (Pixhawk-Cube). Navigating safely through obstacles becomes challenging for UAVs in GPS-denied environments. To address this issue, the drone is configured in altitude hold mode, and the system is trained on the Common Objects in Context (COCO) dataset, enabling it to recognize objects and analyze the surrounding areas to identify free spaces. By doing so, the drone can traverse an obstacle-free path. The detected obstacle information is then utilized to generate avoidance trajectories, allowing the UAV to navigate around obstacles safely. Real-time testing of the proposed technique demonstrates its efficacy in detecting and avoiding obstacles within a threshold distance of 2 meters, with an error rate of 10%. The drone’s relative speed is configured at 2 m/s during these tests.