The purpose of this study is to assess the accuracy and efficiency of the nano versions of the YOLOv5, YOLOv8, YOLOv10, YOLO11 and YOLO12 models in detecting foreign object debris (FOD) at airports and to compare these models using the most commonly used performance criteria to identify their strengths and weaknesses.
The data set used for this purpose is publicly available, referred to as FOD-A (FOD in airports), and it contains 31 object categories and over 30,000 annotation examples. The single runs, conducted to measure the potential of the evaluated models, are followed by a 10-run evaluation process to establish a reliable performance profile.
This study’s findings highlight YOLOv8n’s detection accuracy; however, all models generally perform well in foreign object detection at airports without substantial loss in precision. YOLOv10n, on the other hand, stands out for its performance under various perturbations and on edge computing platforms, as well as its better discrimination of small objects than other models. However, all models have limited capabilities in distinguishing the nail and nut categories and small-sized objects, which should be considered aspects for improvement in model development. In addition to scale- and category-specific challenges, Gaussian noise and image blurring are prominent perturbations that degrade performance.
Real-time object detectors with high performance and low inference times are of interest to both practitioners and researchers in restricted devices. YOLO, a continuously evolving real-time object detector, is still new for foreign object detection research in airports, and there is a need to summarise the available models and make comparisons.
