- The paper presents a novel cross-session fusion approach that integrates 3D LiDAR and monocular camera data to achieve robust, drift-mitigated UAV localization in GPS-denied settings.
- It details a loosely coupled strategy that separates local LiDAR odometry from vision-based global correction, enhancing system simplicity and resilience under extreme conditions.
- Experimental results show that the approach achieves competitive accuracy with minimal sensor setups, outperforming LiDAR-only methods in feature-sparse and low-frame-rate scenarios.
Cross-Session 3D LiDAR and Camera Fusion for UAV Localization in GPS-Denied Environments
Problem Statement and Motivation
Precise UAV localization in GPS-denied settings is essential for applications such as infrastructure inspection and structural health monitoring, where operational environments are characterized by unreliable or obstructed satellite signals. Traditional methods relying solely on cameras, LiDAR, or IMU suffer in visually ambiguous, geometrically degenerate, or low-illumination scenarios, and typically experience significant pose drift without robust, global correction mechanisms. Many existent fusion frameworks require complex sensor setups (stereo pairs, tightly-coupled visual-inertial fusion), limiting simplicity and applicability for small UAVs. The fundamental challenge addressed is robust, accurate, and real-time UAV localization under these constraints, using minimal and lightweight sensor suites.
Method: Cross-Fusion Framework
The paper introduces Cross-Fusion, a hybrid localization architecture leveraging only a 3D LiDAR and a monocular RGB camera. The central technical contributions of the approach include:
- Cross-session fusion: The framework exploits multi-session data integration, aligning online UAV observations with a pre-constructed visual-geometric reference database accumulated from diverse sessions (e.g., ground agents, prior manual UAV flights). This enables leveraging historical information to address environmental aliasing, feature sparsity, and correct pose drift.
- LiDAR-guided 2D-3D correspondence transfer: Visual 2D-2D keypoint matches between online and baseline images are projected into the 3D domain using concurrent LiDAR acquisitions, enabling direct 3D-3D alignment and thus precise 6-DoF pose estimation without reliance on stereo geometry or IMU integration.
- Loosely coupled fusion strategy: The method separates LiDAR-based odometry for local motion tracking from vision-based global correction, balancing simplicity and resilience. Unlike tightly-coupled systems (e.g., FAST-LIVO, LVI-SAM), Cross-Fusion achieves competitive accuracy with reduced complexity and hardware requirements.
The system divides operation into offline and online phases. In the offline stage, a 3D sparse map is constructed via deep CNN feature extraction (SuperPoint/DISK) from images, global retrieval descriptor computation (EigenPlaces), and SfM triangulation (COLMAP). During online operation, LiDAR odometry provides high-frequency local tracking, which is regularly corrected via visual place recognition and 2D-3D correspondence transfer using the constructed baseline. Pose drift is thus mitigated without requiring loop closure in the traditional SLAM sense.
Experimental Evaluation and Results
Experiments were conducted using a quadcopter equipped with a Quanergy M8 LiDAR and Intel RealSense D435 camera. The test environment was a 150 m × 150 m building exterior, with offline surveys providing a sequence of 60 high-resolution images for constructing the initial database.
Key results demonstrate:
- Localization Accuracy: The Cross-Fusion approach achieves localization trajectories closely tracking the GPS ground truth, notably outperforming LiDAR-only odometry which accumulates significant drift over time.
- Drift Mitigation: Quantitative trajectory analyses confirm that the fused approach maintains lower localization error along mission paths compared to LiDAR-only, with the error remaining bounded across keyframes.
- Feature-Sparse and Extreme Condition Robustness: When benchmarked against FAST-LIVO, a state-of-the-art tightly coupled LiDAR-Inertial-Visual SLAM system, Cross-Fusion shows comparable accuracy under nominal conditions. However, under extreme scenarios, such as very low camera frame rates or high-speed flight (causing minimal visual overlap), Cross-Fusion maintains localization integrity where FAST-LIVO fails completely due to the loss of tightly coupled feature correspondences. This highlights operational robustness in hostile or resource-constrained deployments.
- Hardware Simplicity: The system requires only a single monocular camera and a 3D LiDAR, omitting IMUs in the primary fusion, thereby reducing cost, energy consumption, and susceptibility to calibration drift.
Implications and Theoretical Impact
Cross-Fusion broadens the operational envelope of UAV localization, enabling repeatable, accurate, and robust flight in infrastructure-dense environments and other GPS-denied spaces without burdensome sensor integration or extensive online training. This approach strengthens the argument for cross-session data exploitation and supports semi-automated surveying workflows where both UAVs and other agents contribute to baseline mapping. The success of LiDAR-guided 2D-3D transfer also highlights new avenues for combining sparse visual and geometric cues when traditional loop closure is unfeasible, suggesting further generalization to heterogeneous agent networks and multi-session SLAM settings.
From a theoretical perspective, this design demonstrates that loosely coupled, cross-session fusion can approach the performance of more complex, tightly integrated systems, especially when operational constraints prevent rich visual or tightly synchronized inertial acquisitions.
Limitations and Future Directions
Cross-Fusion currently depends on an offline-constructed visual database, limiting applicability in highly dynamic or previously unseen environments. Future research directions will focus on fully online, collaborative multi-agent mapping and on-the-fly database augmentation, as well as adapting the approach to scenarios involving frequent environmental changes or requiring incremental map updates. Additionally, methods for more intelligent candidate selection for visual correction and further acceleration of correspondence matching will enhance scalability for very large environments.
Conclusion
Cross-Fusion presents a robust solution to UAV localization in GPS-denied environments via minimal sensor fusion and efficient cross-session data utilization. It validates the efficacy of combining LiDAR odometry with vision-based global correction from pre-constructed baselines, offering reliability in visually and geometrically challenging settings with a streamlined hardware load. This work underlines the potential for scalable, resilient localization architectures that harness multi-session and multi-agent insights to overcome traditional SLAM limitations in real-world deployment scenarios.
