[Paper Reading] Kimera-Multi: Robust, Distributed, Dense Metric-Semantic SLAM for Multi-Robot Systems
I think this paper has three features:
- Metric-Semantic mesh map
- Distributed loop closure detection
- Robust distributed trajectory estimation
Distributed loop closure detection
- a –global descriptor–> b
- a <–request 3D keypoints and descriptors– b
- a – send 3D keypoints and descriptor –> b
Subsequently, robot β computes putative correspondences by matching the two sets of feature descriptors using nearest neighbor search implemented in OpenCV. From the putative correspondences, robot β attempts to compute the relative transformation using Nistér’s five-point method [79] and Arun’s three-point method [80] combined with RANSAC [56]. Both techniques are implemented in the OpenGV library [81]. If geometric verification succeeds with more than five correspondences, the loop closure is accepted and sent to the robust distributed trajectory estimation module.
GNC 渐近非凸性，用没那么非凸的替代（surrogate）函数序列来作优化，该函数序列会收敛到原来的损失函数。
Robust distributed initialization Between every pair of robots, inlier loop closures (green→) lead to similar estimates for the alignment between frames (green–>). Each outlier loop closure (red→) produces an outlier frame alignment (red–>), which can be rejected with GNC.