FOD Detection using Deep Learning

Abstract

Foreign object debris (FOD) poses serious threat to aviation safety and can cause dam age that costs billions of dollars per annum. Most airports employ manual FOD detec tion methods which are slow and prone to human error. In contrast, automatic FOD detection has significant advantage over manual methods like elimination of human error and continuous monitoring. Compared to conventional object detection, FOD detection is challenging owing to shape variations, background clutter, and primarily small size of objects. In this thesis, we first re-evaluate the existing state-of-the-art ob ject detection algorithms against recently proposed multi-class FOD-A dataset. From empirical analysis, we find that YOLOv5m performs best with 99% detection accuracy on FOD-A test set among all the evaluated models including anchor-based and anchor free object detectors. Due to the fact, we propose multi-class FOD detector based on YOLOv5m. To capture real-world challenges posed by FOD, it is important to learn performance of the model on out of distribution (OOD) data.