Visual-Inertial Mapping with Non-Linear Factor Recovery

Author: Cremers

Year: 2019

Notes:

inertial meas contains a little information after a few second of integration
hierarchical approach: recovers local VIO estimate, then non-linear factors are extracted after marginalization for full BA
VIO layer: fast tracking kp, mapping layer: lighting and invariant k
velocity and biases not estimated: pre integration is not used
performs full BA

VIO:

pyramidal KLT patch tracker with FAST corners
representation of bearing vectors with stereographic projection ie unit length bearing vector parametrized on a 2D plane
IMU error with pre integrated delta with fixed biases
Classic marginalization that leads to a dense prior between all KF pose
VIO pb: Consists of $m$ temporal states (with pose, velocity and IMU bias) and $n$ old keyframes states (only pose) + landmarks

Mapping:

implicit loop detection with ORB feature matching
loop closure with reprojection error and factor recovered by NFR
Here uses NFR to transfer information from VIO to map optimization
optimizes all keypoints + all the factors from NFR
topology: absolute factor on roll-pitch, yaw and absolute position + relative pose