Enhancing Aircraft Safety Through Machine Learning-Based Foreign Object Debris (FOD) Detection

Abstract

In recent years, the issue of foreign object debris at airports has grown significantly. It is noted that airport runways, entryways, and taxiways are the primary locations where foreign object debris-related incidents happen [2]. Foreign object debris can damage the tires or engines of the aircraft, rendering them inoperable. The current foreign object debris detection solutions are limited by their reliance on outdated datasets. To tackle this problem an updated local dataset is used and this dataset was obtained by capturing images through the mobile camera on the airport surface. The dataset consists of more than 30,000 images and 3 categories of foreign objects including concrete (stones), metal (nuts, bolts, nails, etc.), and plastic (bottles, caps, pens, etc.). The proposed framework for identifying foreign object debris integrates augmentation techniques to overcome the difficulties presented by the datasets’ different light levels, noise, and blurry images. Roboflow is used to annotate the images of this dataset. Within this context, the current research introduces an inventive methodology designed to address the issue of foreign object detection, and the newly released You Only Look Once (YOLOv8) with a transfer learning technique. An object detector’s performance is usually assessed, considering inference time and detection accuracy. When compared to two-stage object detectors, single-stage detectors often obtain higher detection accuracy. Optimized parameters are used after the implementation of hyperparameter tuning. The inference time is 4.2ms for 25 epochs. The proposed YOLOv8 model has achieved mean Average Precision (mAP50) of 96.5% and (mAP50-95) of 74.5% for all classes and they improved after using the transfer learning approach. However, a high mAP50 score of 0.975 for the class "metal" in the pre-trained model indicates that the model successfully identifies the instances of metal in the images.