- The paper presents an uncertainty-aware fusion framework that integrates RTK-GNSS and LiDAR data for robust multi-session UAV mapping.
- It employs a coarse-to-fine approach, combining spatiotemporal RTK alignment, scene graph construction, and iterative plane-factor refinement to minimize drift and enhance local geometry.
- Experimental results demonstrate superior global consistency and local precision compared to LiDAR-only methods across diverse environments.
UAV-MapFusion: RTK-Aligned Uncertainty-Aware Coarse-to-Fine Multi-Session UAV Mapping
Problem Statement and Motivation
Large-scale, high-precision 3D point cloud mapping is foundational in robotics, geospatial intelligence, and digital twin systems, particularly for scenarios requiring robust spatial modeling. UAVs, due to their superior mobility and ease of deployment compared to ground vehicles, present an efficient means for point cloud map collection over extensive areas. However, practical constraints like UAV endurance and limited onboard storage necessitate multi-session mapping, where large regions are scanned in several flights. This introduces key challenges for map fusion: significant long-range drift due to error accumulation, sparse or ambiguous inter-session overlap from viewpoint diversity, and non-uniform reliability across heterogeneous sensor measurements. Previous approaches—including place recognition (Scan Context++, BTC, PointNetVLAD) and LiDAR-centric multi-session optimization frameworks (LAMM, MS-Mapping)—lack the mechanisms to simultaneously suppress global drift and guarantee local geometric precision under typical UAV flight dynamics, especially over kilometers and with severe viewpoint changes.
System Architecture and Methodological Contributions
The proposed UAV-MapFusion system is a coarse-to-fine multi-session fusion framework, distinguished by the joint modeling of uncertainty-aware LiDAR and RTK-GNSS observations to achieve globally consistent and locally precise map alignment.
Scene Graph Construction
Intra- and inter-session place association is performed using BTC descriptors, integrating both handcrafted geometric cues and robust triangle-based relations for improved correspondence discovery. Loop closures are geometrically verified via point-to-plane ICP, and a session-level undirected connectivity graph is built. Sessions within the same connected component are merged using RANSAC-based robust pose estimation, providing the initial global alignment in a root-referenced frame.
RTK Spatiotemporal Alignment
RTK observations, while offering valuable global position priors, are often temporally misaligned due to asynchronous sensor streams and susceptible to frame dropouts. UAV-MapFusion introduces a spatiotemporal alignment pipeline:
- Motion excitation detection isolates temporal regions with sufficient dynamic excitation for effective offset estimation.
- Dynamic Time Warping (DTW) robustly estimates the time offset between odometry and RTK sequences.
- A Multi-Output Gaussian Process (MOGP) interpolates RTK positions to synchronize with SLAM odometry timestamps, capturing observation uncertainty and heteroscedastic noise reported by the GNSS receiver.
- A closed-form SVD-based alignment is used to compute the rigid body transformation between the RTK (ENU) frame and the unified map frame.
Uncertainty-Aware Factor Graph Optimization
A factor graph is constructed using GTSAM, integrating odometry, intra- and inter-session loop closures, and RTK positional constraints. All factors are weighted according to observation uncertainties: LiDAR registration residuals, MOGP-predicted covariances for RTK, and plane thickness for planar residuals. The optimization proceeds in two stages:
- Coarse Stage: Corrects accumulated large-scale drift by jointly optimizing all factors.
- Fine Stage: Iterative plane-factor refinement is performed, reconstructing local planar clusters in a hierarchical voxel representation and tightly coupling point-to-plane residuals. This refinement substantially improves structural sharpness and local consistency.
Experimental Results
Evaluations were conducted on four self-developed UAV datasets and the public MARS-LVIG benchmark, covering diverse environments (riverside, forest, factory, farmland, airfield, island, rural, and valley scenes). The test platform used a DJI M300 RTK UAV with an integrated Livox Avia LiDAR and u-blox ZED-F9P GNSS.
Metrics
Map accuracy was quantitatively assessed using Average Wasserstein Distance (AWD) for global consistency and Spatial Consistency Score (SCS) for local map quality, as recommended by MapEval.
Ablation and Benchmarking
- Ablation: Removing the RTK spatiotemporal alignment degraded both AWD and SCS, highlighting the necessity of accurate RTK fusion. The iterative plane-factor refinement reduced residual misalignments and structural blurring, as indicated by further improvements in both metrics.
- Benchmarking: Compared with LAMM and MS-Mapping, UAV-MapFusion consistently achieved lower AWD and SCS across all environments. AW improvements were especially severe in challenging or ambiguous scenes (e.g., forest, islands), demonstrating the benefit of global RTK constraints and local plane optimization. The results quantitatively underscore that LiDAR-only methods are insufficient for kilometer-scale UAV mapping, and that the integration of temporally aligned, uncertainty-weighted multi-source factors is essential.
Theoretical and Practical Implications
The framework demonstrates the effectiveness of integrating RTK priors via continuous-time probabilistic modeling for suppressing global drift, overcoming limitations in classical LiDAR-only SLAM backends. The uncertainty-aware graph structure enables robust fusion under highly heterogeneous measurement qualities, and the iterative plane-factor mechanism improves geometric fidelity in the presence of local misalignments. Importantly, the proposed architecture provides a generalizable template for future multi-sensor, large-scale map fusion in aerial robotics.
Limitations and Prospects
The approach's reliance on planar scene decomposition for local refinement may prove suboptimal in environments with weak planar structure or high irregularity. The current evaluation was with FAST-LIVO2 as the SLAM frontend; future research should validate generality across different odometry sources. Severe RTK data dropouts near session overlaps can still result in imperfect fusion, motivating future work on alternative geometric constraints and more robust optimization in such cases.
Conclusion
UAV-MapFusion delivers a principled, uncertainty-aware, end-to-end method for large-scale UAV point cloud map fusion, addressing the longstanding challenges of multi-session alignment over large and geometrically diverse areas. By bridging the gap between global position priors and local geometric fidelity, this framework achieves state-of-the-art results in both global and local consistency, as evidenced by extensive benchmarking. The public release of code and datasets will facilitate future research in scalable multi-session mapping and robust spatial intelligence systems.
Future Directions
Anticipated developments include extension to scenes with complex non-planar geometries through enriched primitive modeling, validation and tuning for other SLAM frontends, and integration of additional sources of spatial constraints (e.g., semantic or visual cues). The probabilistic multi-sensor alignment methodology may provide the foundation for robust, long-term, multi-UAV collaborative mapping at regional or even city scales.