Author: Stefn Leutenegger
Year: 2022
- Full VI SLAM based on the creation of a pose graph by marginalizing common observation
- BASALT, Kimera, VINS, ORBSLAM are tightly coupled systems = considering visual and inertial measurements with their internal states
- posegraph edges can be turned back into observation reviving old landmarks
- Semantic segmentation with CNN to remove dynamic objects
- OKVIS achieves computationnal tracktability with marginalisation
- factor graph optimization with ceres
- front end: matching, triangulation, seg CNN, place recognition
- realtime estimator: pose graph optimization frame only, sliding window, reviving old observation
- front end with BRISK kp matching
- Keyframe removed based on covisibility criterion
- Marginalization into relative pose factors (?)