A Vision-based Solution for Track Misalignment Detection

Derailments often occur due to railway track misalignments like buckling (lateral) and hogging (vertical). These faults are visually detectable, making automated image-based inspection possible. We introduce the TMD (Track Misalignment Detection) dataset, containing images of normal and misaligned tracks, enabling data-driven detection. The task is formulated as a binary image classification problem using transfer learning (TL). We extract features from pre-trained networks and train lightweight classifiers on them. Since many TL model combinations are possible, we also propose a novel evaluation criterion to identify the most promising models before testing. Experiments show that models selected using our criterion achieve superior performance in detecting track misalignments.