Metrological Assessment of Rail Clip Fasteners via Monocular 3D Reconstruction

Reliable metrological inspection of railway clip fasteners is essential for ensuring correct installation and early detection of deformation, yet it is challenging to achieve with monocular RGB imagery when camera intrinsic parameters are unavailable, and only a limited number of views can be acquired. This work addresses the problem of obtaining repeatable dimensional and angular measurements from such constrained data. We propose an intrinsic parameter-free monocular pipeline that reconstructs clip geometry and performs measurement directly on a depth representation. A neural 3D reconstruction backbone generates a colored point cloud, which is converted into a watertight surface through normal estimation and Screened Poisson meshing. The reconstructed mesh is rasterized to obtain a texture map and a normalized depth map. Clip localization is achieved with a YOLO-based detector trained on the clip head; the region of interest is expanded to include the full clip and segmented using Otsu thresholding on the depth map. Dimensional features are extracted from the segmented depth and mapped to metric units via a spatial sampling factor, while the mounting angle is estimated from characteristic directions associated with the anchored frame and the curved, non-anchored section. Experimental analysis across varying image counts and repeated acquisitions from different viewpoints indicates that the proposed approach supports stable dimensional and angular characterization, and that measurement repeatability improves as reconstruction redundancy increases, demonstrating its suitability for practical inspection with limited RGB inputs.