Authors: Emilio Garcia-Fidalgo, Joan P. Company-Corcoles, Francisco Bonnin-Pascual and Alberto Ortiz
Into review process
- LiDAR only odometry based on point-to-line registering on a local map
- Efficient data structure for map representation
- whole map can be represented as KD-tree but performances degrades with the number of pts ie do not scale well
- Here a hashing scheme is used to represent the map
- Highly inspired by LOAM, inside a factor graph
Related work:
- Loosely-coupled solution with IMU = de skewing point cloud or using IMU to build a prior
- Tightly-coupled: fuse data jointly (in a graph, in a filter..) eg. LIO SAM
- § 3, l9 : a* slinding window
System overview:
- sweep = whole point cloud made of several 2D scans
- odometry module extracts LOAM edges and plane and estimates the current pose with registration on a adaptive local map
- mapping build the global and local map
LiDAR Odometry:
- two thread
- first thread: feature extraction with curvature score
$c$ (the formula can be reminded) - sectors division + non max suppression for edge points
- second thread: pose optimization
- point to line distance computed with the nearest neighbours of an edge point in the local map
- weight on point to line factor based on range of the lidar meas.
- Pas mal, assez clair
LiDAR Mapping:
- Maps of cells, that are cuboids of fixed sizes
$M = { \mathbb{H}, \mathbb{C} }$ $\mathbb{H}$ is the hash table and$\mathbb{C}$ is the set of cells - Map of only edge points=> more sparse
- Local map built with the cells up to a certain range of the current robot pose, it is thus adaptive
PAPER REJECTED: The contribution of the paper is too small: a voxel-based map and point-line features. Feature robustness and characteristics should be detailed as it is part of the contribution. It is not proven that this new mapping gestion is more efficient than KD Trees