Análise comparativa de algoritmos de detecção de objetos para inspeção de trens de pouso

Abstract

Aircraft landing gear visual inspection is a critical stage in Maintenance, Repair, and Operations (MRO) processes, traditionally performed manually and subject to human inconsistencies. This work proposes and comparatively evaluates two architectural paradigms for object detection aimed at automating this task: YOLO11s, based on Convolutional Neural Networks (CNN), and RT-DETR-L, based on Vision Transformers. Both models were trained under rigorously equalized experimental conditions — 50 epochs, the same hardware platform (NVIDIA Tesla T4 GPU), the same dataset (Landing-Gear IRP, 2,979 images with data augmentation), and the same default hyperparameters — using Transfer Learning from weights pre-trained on COCO. The results demonstrated the technical feasibility of both paradigms, with performances superior to those reported in related work: YOLO11s achieved an mAP@50 of 79.03% (precision: 81.35%, recall: 75.18%) and an inference speed of 62.5 FPS, while RT-DETR-L recorded an mAP@50 of 75.94% with a higher recall of 78.50%, a maximum F1 of 0.80, and 18.7 FPS. The comparative analysis revealed well-defined trade-offs: YOLO11s presents higher localization precision (mAP@50-95: 37.76% vs 34.04%) and operates above the 30 FPS threshold required for real-time video, making it recommended for autonomous systems with hardware constraints; RT-DETR-L reduced the false negative rate from 17.3% to 12.0% (30 additional instances detected), making it preferable for inspection-assistance systems with human validation, where minimizing undetected components is a priority.