This study presents the development and performance evaluation of a real-time obstacle recognition system for unmanned aerial vehicles (UAVs) in indoor environments. Utilizing the YOLOv5 object detection model, the system processes video streams from onboard cameras to detect obstacles under varying lighting conditions and distances. Experiments using an F450 UAV with analog (1200VTL) and digital HD cameras measured detection accuracy, recall, F1-score, and frame rate (FPS). The HD camera-based system achieved an average F1-score of 81.4% and over 24 FPS, confirming reliable real-time detection. In low-light or poor image quality conditions, detection accuracy decreased, emphasizing the importance of hardware and video quality. This study experimentally verifies vision-based obstacle detection for indoor UAV navigation and offers a foundation for future integration with autonomous avoidance algorithms.
Park et al. (Tue,) studied this question.