LiDAR-based Registration against Georeferenced Models for Globally Consistent Allocentric Maps

TL;DR

The paper addresses GNSS unreliability in urban environments by registering small LiDAR local maps to georeferenced 3D models built from CityGML and DEM data. It introduces a ray-tracing based plausibility score on a 2D height map to select plausible LiDAR-model alignments and integrates these refinements into a continuous-time spline-based pose graph that fuses LiDAR odometry, GNSS, and IMU constraints to produce globally consistent, georeferenced trajectories and maps. Key contributions include (1) a GNSS refinement strategy via LiDAR-model registration, (2) an occupancy-based plausibility check for local alignments, and (3) a spline-based allocentric optimization framework that anchors the global map to prior geospatial models. Experiments on two testing sites show the method reduces GNSS offsets from up to 16 meters to under 0.5 meters and yields globally consistent maps aligned with prior 3D geospatial models, improving fusion for SAR operations in challenging urban canyons.

Abstract

Modern unmanned aerial vehicles (UAVs) are irreplaceable in search and rescue (SAR) missions to obtain a situational overview or provide closeups without endangering personnel. However, UAVs heavily rely on global navigation satellite system (GNSS) for localization which works well in open spaces, but the precision drastically degrades in the vicinity of buildings. These inaccuracies hinder aggregation of diverse data from multiple sources in a unified georeferenced frame for SAR operators. In contrast, CityGML models provide approximate building shapes with accurate georeferenced poses. Besides, LiDAR works best in the vicinity of 3D structures. Hence, we refine coarse GNSS measurements by registering LiDAR maps against CityGML and digital elevation map (DEM) models as a prior for allocentric mapping. An intuitive plausibility score selects the best hypothesis based on occupancy using a 2D height map. Afterwards, we integrate the registration results in a continuous-time spline-based pose graph optimizer with LiDAR odometry and further sensing modalities to obtain globally consistent, georeferenced trajectories and maps. We evaluate the viability of our approach on multiple flights captured at two distinct testing sites. Our method successfully reduced GNSS offset errors from up-to 16 m to below 0.5 m on multiple flights. Furthermore, we obtain globally consistent maps w.r.t. prior 3D geospatial models.

Figures (8)

• Figure 1: Geospatial maps [a)] contain approximate building shapes and ground height. Inaccurate raw gnss measurements impair the accuracy of georeferenced maps [b)]. We obtain a globally consistent map [c)] by registration against the geospatial model.
• Figure 2: System overview: Our refinement aligns small LiDAR maps against a geospatial model using gnss for initialization. After pose graph optimization, our system outputs a globally consistent and georeferenced map and trajectory.
• Figure 3: Model Generation: A view [a)] on the Poppelsdorf campus at the University of Bonn for easier scene understanding. The cgml data [b)] contains the rough polygonal building shape, while the DEM [c)] represents the ground surfaces. We combine the subsampled polygonal cgml [d)] with the interpolated DEM [e)] for our model [f)]. A surfel map [g)] is derived for registration quenzel2021mars and a height map [h)] for our plausibility check.
• Figure 4: A semantically annotated LiDAR scan with vegetation (green) and people (yellow) before [a)] and after filtering [b)].
• Figure 5: Spline-based Pose Graph: We estimate a continuous-time B-spline trajectory sommer2020cvpr$T(t)$ with $N$ knots ($\bm{x}_i,\ldots,\bm{x}_{i+N-1} \in \mathcal{X}$) being active per scan. Raw and refined gnss positions ($T_\mathrm{gnss},T_\mathrm{ref}$) allow to georeference the UAV trajectory with an anchor pose $T_\mathrm{a}$. Odometry constraints ($e_\mathrm{o}$) connect each scan $\mathcal{P}$ with the previous keyframe. Preintegrated IMU measurements ($e_\mathrm{\Delta}$) with optimizable biases ($\bm{\omega}_\mathrm{b},\bm{\alpha}_\mathrm{b}$) enforce smoothness within a scan. Relative poses ($e_\mathrm{r}$) between keyframes or scans enable loop closing.