Author: Yi
Year: 2023
- LiDAR SLAM for computation limited platforms
- Select non-conspicuous (non-visibles) features, then two stage matching (KD-tree + graph matching) that gives weight for odometry optim
- two streams for scan registration:
- ICP
- feature extraction (LOAM)
- Remove disjoint points, i.e. points at the borders that may represent outliers
- Then compute the smoothness score to extract corners and planes
- divise each laser beam into 6 subregions and select weak corners and planes in each subregions as optimized features
- KD tree matching may lead to multi to one coresp case
- geometric consistency: the euclidean distance between targeted points remains constant through scans
- build a compatibility graph based on each hypothetical association and remove edges based on geo const
- point to line residual and point to plane residual weighted by association scores from the graph matching
- mapping with two stage associations + ptl and ptp optim
- experiments on KITTI and on a 32 beam lidar embedded on a quadrotor