COMPARATIVE BENCHMARKING OF YOLO11 FOR UAV-BASED HUMAN DETECTION IN SIMULATED DISASTER SCENARIOS
TL;DR
YOLO11m provides a promising accuracy–efficiency baseline for UAV-based SAR research, but because the evaluation relies on synthetic data and does not include onboard inference or physical UAV field trials, further validation using real-world aerial disaster imagery is required before operational deployment.
Abstract
Search and Rescue (SAR) operations require rapid and reliable identification of survivors in disaster-affected areas. This study presents an empirical benchmark of YOLO11m for UAV-oriented human detection in simulated disaster scenarios. The model was trained and evaluated using the synthetically generated C2A: Human Detection in Disaster Scenarios dataset and compared with YOLOv8m, YOLOv9m, and YOLOv10m under consistent training configurations. The experiments were conducted at a resolution of 640 × 640 pixels using mixed-precision training on dual NVIDIA T4 GPUs. YOLO11m achieved an mAP@50 of 0.850, an mAP@50–95 of 0.612, a Precision of 0.880, a Recall of 0.799, and a peak F1-score of 0.830 at a confidence threshold of 0.376. The model also recorded an average inference latency of 5.5 ms per image in the test environment. Compared with the evaluated medium-scale YOLO variants, YOLO11m achieved the highest mAP@50–95 while maintaining moderate parameter and computational requirements. These results indicate that YOLO11m provides a promising accuracy–efficiency baseline for UAV-based SAR research. However, because the evaluation relies on synthetic data and does not include onboard inference or physical UAV field trials, further validation using real-world aerial disaster imagery is required before operational deployment.