Search

Automated Foreign Object Debris (FOD) detection offers significant benefit to the aviation industry by reducing human error and enabling continuous surveillance. This paper focuses on addressing the intricacies of FOD detection, with a specific emphasis on treating FODs as "small" objects, a facet which has received limited attention in prior research. This study provides a pioneering evaluation of state-of-the-art object detectors, including both anchor-based models including SSD, YOLOv5m, Scaled YOLOv4 and anchorless models CenterNet and YOLOv8m, applied to a multiclass FOD dataset, "FOD in Airports (FOD-A)", as well as meticulously curated subset of FOD-A featuring small FODs. The findings reveal that the anchorless object detector YOLOv8m gives the best time accuracy trade off outperforming all compared anchor-based and anchorless approaches. To address the challenge of detecting small FODs, this study optimizes YOLOv8m model by making architectural modifications and incorporating a dedicated, shallow detection head that is purpose-built for the precise identification of small objects. The proposed model, termed as "Improved YOLOv8", outperforms YOLOv8m by a margin of 1.02 in Average Precision for small objects (APs), achieving a mean average precision (mAP) of 93.8%. Notably, Improved YOLOv8 also has better mAP than all the considered anchor-based and anchorless object detectors examined, as well as those featured in prior FOD-A dataset research. [ABSTRACT FROM AUTHOR]