Rail Fastener Defect Inspection Based on UAV Images: A Comparative Study

Abstract

Rail fastener seriously affects the safety of railway system. The main approaches for detection of rail fastener defect are human inspection and rail vehicle inspection, which have many drawbacks such as low efficiency, high cost and so on. The paper presents a novel UAV-based visual inspection mean for fastener defect and focuses on two kinds of approaches based on UAV rail images: methods using traditional visual and using deep learning. With regards to the first aspect, a new traditional detection method using SVM and HOG features is proposed, which acquires a low mAP, due to heterogeneous background, various illumination, small target and so on. With regards to the second, some approaches based on deep learning including Fastener RCNN, YOLOv3, improved YOLOv3 and FPN are compared and applied for fastener defect detection, which can achieve a satisfied result. Finally, a quantitative comparison experiment shows FPN acquires a mAP of 95.78%, which achieves favorable performance against state-of-the-art methods. Therefore, it is possible to carry out fastener defect inspection with UAV images.