Low-Drift Odometry, Mapping and Ground Segmentation Using a Backpack LiDAR System

Robust Multi-Modal Multi-LiDAR-Inertial Odometry and Mapping for Indoor Environments

“…To merge clouds into one combined cloud P m , we need to align the starting timestamp and ending timestamp. We adopt a splitand-merge method similar to [13].The individual timestamp of p h i ∈ P h and p v i ∈ P v can be obtained from the sensors driver. If the timestamp for a point p v i ∈ P v is not available, it also can be calculated by orientation difference [18].…”

“…If the depth difference between the point in L v i or L h j and nearest neighbor points within the same subset is smaller than the depth threshold d th , then the point is added to continuous points subset P h i C or P v j C . Then we follow the feature extraction methods in [13], where a scatter matrix Σ is calculated based on neighbor points. By analyzing the two largest eigenvalues λ 1 and λ 2 of Σ, the plane points are detected and labeled as a plane, the point where two plane meet is labeled as corner features, the points where one plane ends and neighboring with discontinuous points are labeled as break points.…”

“…Multiple studies in the literature have focused on improving LiDAR maps by integrating point clouds from multiple LiDAR sensors [12], [13]. However, the low frame publishing frequency of typical LiDARs (e.g., 10 Hz) can hinder an accurate 6-DOF pose estimation in multi-LiDAR systems.…”

SLICT: Multi-input Multi-scale Surfel-Based Lidar-Inertial Continuous-Time Odometry and Mapping

“…Since lidar input is acquired at 100ms, currently real-time performance is not guaranteed by SLICT. Because we associate the points with surfels at five scales (from 2 1 to 2 5 , where = 0.1m), the computation load for association is at least a multitude that of direct method, which uses only one voxel scale. However, we think it is justifiable considering that SLICT gives higher accuracy, and real-time performance can be achieved by using a CPU that supports more threads.…”

“…However, as the map grows, recomputing the k-d tree on the global map is intractable. One strategy to avoid this issue is to organize the maps into keyframes and build a local map with bounded size from a finite number of nearest keyframes [3]- [5], [7], [14]. However this strategy is sub-optimal since there are cases where the local map misses important prior observations that were captured in previous keyframes (an example is given in Fig.…”