Abstract
Railway wagon inspection is a key enabler of digital railway operations and intelligent asset management that requires a complete view of each wagon side, which a fixed trackside area-scan camera cannot capture in a single frame as the train passes. We present a complete image stitching pipeline that reconstructs each wagon from its sequence of overlapping frames under uncontrolled outdoor conditions: day and night illumination, varied wagon surfaces, background clutter and camera vibration. The pipeline applies seven sequential keypoint filters to dense correspondences produced by a learned dense matcher, followed by robust median displacement estimation and a sequence-level outlier recovery mechanism. We also propose a three-tier evaluation framework that scores a stitching system from its reported geometry alone, without access to its internals, supported by a benchmark dataset of 30 railway wagons across six trains with 262 manually annotated frame pairs. End-to-end evaluation on the benchmark dataset shows that the pipeline recovers accurate displacements where the raw correspondences yield none; no single filter dominates on every measure, with the decisive filter depending on the surface; and the complete system outperforms both a classical sparse-feature baseline and a generic dense-matching baseline. The evaluation framework is publicly released, and the annotated dataset is available on request, to facilitate reproducibility.