A New Visual Inertial Simultaneous Localization and Mapping (SLAM) Algorithm Based on Point and Line Features
Author: Zhang
Year: 2022
- bilateral filtering -> surf extraction -> FLANN
- comparison to PL-VIO visu-inertial method basde on point line features
- robust to illumination and blur
- point based perform hardly in low texture envi like corridors => add line features
- Line Segement Detector (LSD) adjusted to meet real time
- LKT doesn't work in env with light changes => FLANN
Contributions:
- adaptive line segment algorithm for line processing
- bilateral matching
- visual inertial initialization
Front end:
- adaptive filtering: blurs areas without texture, enhances edges
- point -> surf -> flann -> KD tree for outliers
- line -> LSD -> geometric constraint for matching -> RANSAC
- FLANN: we take the smallest and the second smallest euclidean distance between features and if the ratio between the two is smaller than a thresh it is a match
- parameter tunning for LSD + a heuristic to determine if a segment is good
- geometrical constraint to match lines : lines are matched if they minimize angle, length ration, projection ratio and midpoint flow (procédé intéressant pour prendre en compte différents critères de matching)
IMU intialization:
- scale, bias and gravity vector are estimated on several KF
- bias are updated through time