From Benchmarking to Optimisation: A Comprehensive Study of Aircraft Component Segmentation for Apron Safety Using YOLOv8-Seg

Abstract

Apron incidents remain a critical safety concern in aviation, yet progress in vision-based surveillance has been limited by the lack of open-source datasets with detailed aircraft component annotations and systematic benchmarks. This study addresses these limitations through three contributions. First, a novel hybrid dataset was developed, integrating real and synthetic imagery with pixel-level labels for aircraft, fuselage, wings, tail, and nose. This publicly available resource fills a longstanding gap, reducing reliance on proprietary datasets. Second, the dataset was used to benchmark twelve advanced object detection and segmentation models, including You Only Look Once (YOLO) variants, two-stage detectors, and Transformer-based approaches, evaluated using mean Average Precision (mAP), Precision, Recall, and inference speed (FPS). Results revealed that YOLOv9 delivered the highest bounding box accuracy, whereas YOLOv8-Seg outperformed in segmentation, surpassing some of its newer successors and showing that architectural advancements do not always equate to superiority. Third, YOLOv8-Seg was systematically optimised through an eight-step ablation study, integrating optimisation strategies across loss design, computational efficiency, and data processing. The optimised model achieved an 8.04-point improvement in mAP@0.5:0.95 compared to the baseline and demonstrated enhanced robustness under challenging conditions. Overall, these contributions provide a reliable foundation for future vision-based apron monitoring and collision risk prevention systems.