VIGS-Fusion: Fast Gaussian Splatting SLAM processed onboard a small quadrirotor
Résumé
Real-time dense visual-inertial SLAM remains a major challenge for resource-constrained UAVs, especially when using 3D Gaussian Splatting (3DGS). While 3DGS enables high quality 3D reconstruction, it suffers from high computational cost and weak tracking robustness. In this paper, we introduce VIGS-Fusion, an efficient 3DGS SLAM system that can operate fully onboard. Our method computes a robust and accurate pose from a single tightly-coupled optimization problem that incorporates RGB-D data, preintegrated IMU measurements, velocity, and IMU bias. By using a second-order approximation and a loss based on the observed image gradients for pose estimation, we significantly reduced computational time while maintaining high tracking performance. We also provide a real-world dataset collected with our compact quadrotor, capturing RGB-D and IMU measurements. Experimental results demonstrated accurate and robust tracking while achieving over a 30x reduction in computation time for the tracking compared to state-of-the-art 3DGS SLAM systems. For the first time, we showed real-time 3D Gaussian Splatting SLAM processed entirely onboard a 6-inch UAV. The code is available at: https://github.com/AbdoullahNdoye/VIGS-Fusion.git.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |