Deep learning approach for foreign object detection on airport runways

The aim of this thesis is to establish a deep learning-based object detection system for identifying foreign object debris (FOD) on airport runways and to conduct a comparative performance evaluation of different YOLO architectures (YOLOv5, YOLOv8, and YOLOv11). Foreign objects on runways pose a serious threat to flight safety, making their fast, accurate, and reliable detection critically important. The central hypothesis of the thesis is that modern deep learning-based object detection algorithms offer higher accuracy, lower false negative rates, and greater robustness to varying environmental conditions compared to traditional methods for FOD detection. Accordingly, comprehensive experiments were conducted using publicly available YOLOv5s, YOLOv8s, and the latest YOLOv11-s models. The dataset used in this study is the FOD-A dataset, which was collected in airport environments and includes 31 different FOD classes under various weather and lighting conditions. The models were evaluated using performance metrics such as precision, recall, F1-score, mAP@0.50, and mAP@0.50–0.95. According to the results, the YOLOv11-s model, despite its lightweight architecture, outperformed YOLOv5s and YOLOv8s in terms of F1-score (0.861) and recall (0.826). This demonstrates a significant advantage in minimizing false negatives, especially in safety-critical applications. While the YOLOv8s model stood out in terms of overall precision and mAP performance, YOLOv5s exhibited lower recall in certain scenarios. Furthermore, the YOLOv11-s model, with only 9.45 million parameters and a training time of approximately 45 minutes on a mid-range GPU, offers high applicability in resource-constrained embedded systems or real-time airport applications. In conclusion, this thesis does not introduce a new architecture but provides an in-depth analysis of the performance of existing YOLO models. The findings highlight that deep learning methods offer significant advantages over traditional systems in FOD detection, and the YOLOv11-s model, in particular, emerges as an effective and scalable solution for real-time runway safety due to its lightweight structure and strong generalization capability.