A graph-matching approach for cross-view registration of over-view and street-view based point clouds

Wide-area 3D data generation for complex urban environments often needs to leverage a mixed use of collected from both air and ground platforms, such as aerial surveys, satellite, mobile vehicles. On one hand, kind with information drastically different views (ca. 90° more) forming cross-view data, which due very limited overlapping region caused by the line sight sensors, is difficult be registered without significant manual efforts. other registration suffers non-rigid distortion street-view (e.g., trajectory drift), cannot simply rectified similarity transformation. In this paper, based on assumption that object boundaries buildings) over-view should coincide footprints façade points generated photogrammetric images, we aim address problem proposing fully automated geo-registration method utilizes semantically segmented view-invariant features under global optimization framework through graph-matching: taking point clouds stereo/multi-stereo satellite images monocular video inputs, proposed models segments buildings nodes graphs, detected satellite-based clouds, thus form graph-matching allow matches; enable robust solution utilize topological relations between these segments, propose its conjugate graph solved belief-propagation algorithm. The matched will subject further precise-registration, followed constrained bundle adjustment image keep 2D-3D consistencies, yields well-registered clouds. Our assumes no or little prior pose (e.g. sparse locations consumer-grade GPS (global positioning system)) has been applied large dataset scale difference containing 0.5 m GSD (Ground Sampling Distance) 0.005 1.5 km in length involving 12 GB data. experiment shows achieved promising results (1.27 accuracy 3D), evaluated using LiDAR Furthermore, included additional experiments demonstrate can generalized process types sources, e.g., open street view maps semantic labeling maps. Codes made available Github Repository.1