Author: Ferrera & co
Year: 2021
- efficient multi-threaded architecture (4 threads)
- when ORBSLAM is RT forced, accuracy decreases
- online bag of word for loop closing (first time in SLAM)
- PTAM first BA method in real time thx to parallelization
- limits triangulation by matching the most to existing points in the map
- CLAHE = contrast enhancement
- pyramidal LKT for keypoint tracking (if 3D keypoints 2 levels, if 2D kp 4 levels) with reverse tracking
- RANSAC with 5 points to estimate the essential matrix with 3D points only
- initial guess with constant velocity assumption and then P3P RANSAC
- outlier rejection after reprojection error
- Keyframe creation based on parallax
- LKT for stereo matching
- retracking steps on the local map can be considered as small loop closures
- tracking on local map with 3D position of keypoints and BRIEF descriptors
- include in BA every KF that have at least 25 common observations
- 3D map points are parametrized with 2D pixel pose on the first KF and its inverse depth => one DOF per 3D point
- LC uses both keypoints from the SLAM and additionnal FAST features with BRIEF descriptors
- Pose graph optimization to take LC into account, then for each kp its new world pose is computed as
$\lambda^{w^*} = {}^w T_{\alpha}^{new} {}^{\alpha}T_w^{old} \lambda^w$ - 200Hz SLAM by disabling LC and use of FAST features with a 50*50 pixel grid for KF creation
Très bon papier avec plein de petites astuces à réutiliser pour mon projet