Performance Analysis of Edge AI Models for Object Detection in Unmanned Aerial Vehicles (UAVs): A Comparative Study of YOLOv8n, YOLOv9c, and RT-DETR-l

Real-time object detection from bird's-eye view perspectives in Unmanned Aerial Vehicles (UAVs) and Edge AI systems remains a significant engineering challenge due to small object scales, complex backgrounds, and limited computational resources. In this study, the performance of convolutional neural network (CNN)-based YOLOv8n and YOLOv9c models, alongside the Vision Transformer (ViT)-based RT-DETR-l model, is experimentally evaluated on the VisDrone2019 dataset. To ensure a fair comparison, all models were trained in an NVIDIA Tesla T4 GPU environment and evaluated at the 45th epoch checkpoint. Experimental results demonstrate that the RT-DETR-l model achieves the highest detection accuracy with a 0.492 mAP50, albeit at a higher computational cost with a 20.9 ms inference latency. Conversely, YOLOv9c provides a balanced trade-off with a 0.452 mAP50 accuracy and a 9.6 ms latency, whereas YOLOv8n stands out as the lightest solution for constrained embedded hardware, achieving 1.5 ms latency and 0.313 mAP50 accuracy. Based on these findings, model selection criteria are outlined for various operational UAV mission profiles, such as SAR (Search and Rescue), wildfire detection, and autonomous patrols.