A Tightly Coupled Feature-based Visual-Inertial Odometry with Stereo Cameras

Author: Lei Yu

Year: 2022

Notes:

tightly coupled in the backend and in the front end
IMU aided feature descriptors
fast feature tracking based on IMU
feature based methods with descriptor matching are more frequently applied in engineering
ORB do not describe in plane (roll) and struggle with out of plane (pitch, yaw) rotations

IMU aided front end:

IMU predict the camera pose by integratio of $a$ and $\omega$
perspective deformation due to camera motion
uses a perspective deformation of the image patch to build the descriptor
prediction of the pose of a corner in the next frame with IMU, then performe descriptor matching with all the kp in a radius of 10 around

Backend:

First pose is refined with EPNP Ransac
CERES with pre integration, 3D to 2D residuals and 2D to 2D residuals

Results:

~ 75 ms per iteration
IMU aided descriptor better than BRIEF or ORB on the trajectory precision level
better than VINS FUSION & co on every EUROC sequence

Papier de grande qualité