Evaluating YOLO Variants With Transfer Learning for Real-Time UAV Obstacle Detection in Simulated Forest Environments

Autonomous navigation in forest environments poses significant challenges for Unmanned Aerial Vehicles (UAVs), where reliable obstacle detection is essential for safe operation. This study presents a detailed performance evaluation of recent YOLO models (v8–v11) using a realistic simulation environment built on Unreal Engine and AirSim. To the best of our knowledge, this is the first study to comprehensively benchmark YOLOv8–v11 in a UAV-based obstacle detection context, incorporating both onboard deployment factors and the impact of transfer learning (TL) using EfficientNet backbones. We analyze detection accuracy (mAP@50, precision, recall), inference time, FPS, and memory usage across all YOLO configurations, from lightweight (YOLOv8s) to high-performance models (YOLOv11x). TL significantly improved model performance; for instance, YOLOv8s showed a recall increase from 0.6318 to 0.7804 and mAP@50 from 0.7228 to 0.8447. YOLOv8m and YOLOv10b emerged as optimal models, offering the best trade-off between speed and detection accuracy for real-time UAV deployment. Larger models like YOLOv9c and YOLOv10x delivered superior accuracy but required more computational resources. The results offer practical guidance for selecting YOLO architectures in UAV-based obstacle detection tasks, balancing accuracy, speed, and deployment feasibility in complex environments.