TrackSafe: a comparative study of data-driven techniques for automated railway track fault detection using image datasets

Railway track accidents continue to occur despite manual inspections, which are often inaccurate and can lead to catastrophic events. While artificial intelligence has been applied in the railway sector, few studies have focused on defect detection using object detection tools. Additionally, there is a lack of studies that compare different models using the same dataset.

This paper proposes new data-driven techniques that identify railway track faults using three object detection models: YOLOv5, Faster RCNN, and EfficientDet. These models are compared by testing a dataset of 31 images that contain three different railway track elements (clip, rail, and fishplate), both faulty and non-faulty. Six classes were differentiated in the training of the models: one faulty and one non-faulty for each of the three classes. Image pre-processing steps included data augmentation techniques and image resizing. Results show good precision (equivalent to 1) in detecting non-defective elements, but recall values for defective elements vary among models, with Faster RCNN performing the best (0.93), followed by EfficientDet (0.81), and YOLOv5 (0.68). The full paper discusses the strengths and weaknesses of these proposed techniques for railway fault detection.