Real-Time Runway FOD Detection with YOLO26s

Abstract

The safety of runways and industrial facilities in general depends on the rapid detection of foreign object debris (FOD). This study proposes to use artificial intelligence models with the help of unmanned aerial vehicles (UAVs), namely You Only Look Once 26 (YOLO26s). Because of changes in the architecture, such as using the new Muon Stochastic Gradient Descent (MuSGD) optimizer, deleting the old loss functions, and using the new progressive loss function (ProgLoss), this model is ideally suited for real-time use at low power units, for example, companion computers. The results of testing on a set of 1464 images of real FOD objects (tools, parts) confirmed its high reliability: the mean Average Precision (mAP50) metric reached 0.983, Precision - 0.947, and Recall - 0.976. The model weight is 20.3 MB. The model demonstrated an exceptional card processing speed of 97 frames per second (FPS), which makes this model very effective for real-time operation and for FOD detection.