The comparison of distance metrics in descriptor matching methods utilised in TLS-SfM point cloud registration

Advanced measurement techniques, such as Terrestrial Laser Scanning (TLS), play a vital role in documenting cultural heritage and civil engineering structures. A key aspect of these applications is the accurate registration of point clouds. Conventional TLS methods often rely on manual or semi-automated correspondence detection, which can be inefficient for large or complex objects. Structure-from-Motion Terrestrial Laser Scanning (SfM-TLS) offers an alternative methodology, comprising two primary phases: correspondence search and incremental reconstruction. Descriptor matching in SfM-TLS typically employs the L2 norm to measure Euclidean distances between features, valued for its simplicity and compatibility with algorithms like SIFT. This study investigates the influence of various distance metrics on descriptor matching during the correspondence search stage of SfM-TLS. Eight metrics were analysed: Bray-Curtis, Canberra, Correlation, Cosine, L1, L2, Squared Euclidean, and Standardised Euclidean. Synthetic data experiments highlighted challenges in keypoint detection and matching due to measurement angles, material characteristics, and 3D-to-2D transformations. Simulations incorporating Gaussian noise demonstrated that image rotation and skew significantly affected tie-point accuracy, more so than variations in intensity. In field applications involving cultural heritage sites and building interiors, the L1 and Squared Euclidean metrics yielded higher accuracy, while the Canberra metric underperformed. Metric selection was found to have a greater impact on complex geometries, such as historical structures, compared to simpler forms. Consequently, this study recommends the L1 and Squared Euclidean metrics for pairwise SfM-TLS registration, as they exhibit robustness in maintaining high accuracy and completeness across a variety of architectural scenarios.