Abstract
Monocular visual-inertial simultaneous localization and mapping (SLAM) estimates the 3D positions of landmarks through triangulation of inter-frame tracked features. However, due to limited parallax between frames and highly consistent camera motion, the triangulation process frequently becomes ill-conditioned. To address this, we propose MR2SLAM, a Multi-Robot Monocular Resilient SLAM for ill-conditioned problems. MR2SLAM leverages collective observations of feature points from multiple robots to perform joint triangulation, benefiting from diverse viewpoints, complex epipolar geometry, and improved robustness against ill-conditioning. Furthermore, to fully leverage multi-robot collaborative observations and mitigate the impact of visual data degradation in any single robot on the overall system, MR2SLAM resiliently adjusts the weights between map points generated through multi-robot collaboration and those from individual robots in back-end optimization. This is achieved by updating the information matrix using reprojection errors. We evaluate our method on a multi-vehicle dataset collected in underground environments. Results show that, compared to VINS-Mono, the proposed method reduces the average condition number by 92.56% and improves the average localization accuracy by 19.68%.
Funding Information
  • National Natural Science Foundation of China (42330113)

This publication has 18 references indexed in Scilit: