Robust Cross-Modal Matching between LiDAR Point Clouds and Multi-Camera Images in Tunnel Environments via Surface Parameterization

Abstract. Tunnel lining defect detection increasingly depends on LiDAR and multi-camera collaboration for high-precision 3D modeling. However, sparse textures, repetitive structures, and insufficient lighting limit camera overlap, compromising target-free cross-modal matching. We propose leveraging inherent tunnel structural regularity for robust matching by projecting LiDAR intensity data and stitched images onto a shared parameterized surface. Leveraging geometric priors, LiDAR point clouds are fitted to cross-sections and unrolled into a 2D intensity map within a longitudinal-angular domain. Simultaneously, eight images undergo photometric normalization and stitching into a panoramic image sampled in the same domain. Feature extraction and matching occur in this unified 2D space, with correspondences mapped back to 3D-2D relationships via invertible bi-directional mapping. This approach facilitates geometric constraints and consistency checks to filter unreliable matches. Experiments on the Daliang and Zhaobishan railway tunnels demonstrate that our method generates significantly more geometrically verified correspondences compared to direct image matching. Matching stability is enhanced, with high alignment precision at structural edges, effectively supporting subsequent fusion and analysis tasks.