Investigation of different registration methods for TLS-based deformation analysis of hydroelectric dams – A case study

Abstract

In this paper, we investigate various registration techniques for Terrestrial Laser Scanning (TLS) in the context of deformation analysis of hydroelectric dams. Accurate spatiotemporal registration of TLS data is particularly challenging in non-urban and mountainous environments due to the scarcity of unobstructed and geometrically well-defined surfaces. This is compounded by the presence of unknown changes over time in potentially large parts of the scanned scenes. These challenges complicate the establishment of suitable correspondences between the scans. Traditional registration methods often struggle under these conditions, leading to point cloud differences that may be misinterpreted and mask the actual deformations. We apply an approach utilizing optical flow, as well as Feature to Feature Supervoxel-based Spatial Smoothing (F2S3), to determine 3D vector fields between corresponding points and robustly estimate the registration parameters from these correspondences. We conduct a comparative analysis of the registration accuracies achieved using the above methods and those obtained from traditional registration methods, including the Iterative Closest Point (ICP) algorithm. Target-based registration results serve as a benchmark for this analysis. Additionally, we study the impact of the various registration approaches on the estimated deformations and compare the TLS-based results to those obtained from plumb line measurements within the dam. The presented investigation uses real measurements from the Santa Maria dam in the Swiss Alps, but the findings are transferable to other geomonitoring application cases in non-urban environments.