Author: Carlos Campos
Year: 2021
- V, VI and multi map slam with pinhole, rgbd and fisheye lens model
- multi map used for relocalization
- Key of SLAM accuracy: long term data association
- improved place recognition by checking geometric consistency first and then local consistency
- Atlas = multi map system
- abstract camera representation => test new camera models :(
- Direct slam map estimation is reduced to pose map => less accuracy than sparse SLAM
- VINS Fusion KLT with shi tomasi kp
- ORB SLAM VI initialization too slow
GREAT comparison between state of the art SLAM systems
Camera model:
- All properties and function depending on the camera model are extracted
- MLPnp (maximum likelihood pnp) algorithm decoupled from the camera model => uses projective ray as input
- do not rely on stereo rectification => bo monocular stereo system
Visual Inertial SLAM:
- IMU pre integration on manifold
- Keyframe based VI SLAM
- huber loss not needed for inertial observations
- Inertial initialization stated as a MAP estimatio problem of: the scale, the orientation of g, the bias and the velocity on 10 KF taken in 2 seconds
- In some specific cases, when slow motion does not provide good observability of the inertial parameters, initialization may fail to converge to accurate solutions in just 15 seconds.
Loop closing:
- DBOW2 can achieve 100% precision but 40% recall (=proportion d'items pertinents proposés sur l'ensemble d'items pertinents)
- Atlas system: active and matching map merging + loop closure with current KF and active map
- For each Loop closure hypohtesis, a local window of covisible KF on the candidate is selected to proceed to 3D alignement on a matching map and to validate or not the loop
- map merging operation: remove rendundant points between M_a and M_m, peroform local BA on the merged map with camera pose fixed and then pose graph optimization is performed
- loop closure = map merging but where both KF belong to active map