LF-VIO: A Visual-Inertial-Odometry Framework for Large Field-of-View Cameras with Negative Plane

Year: 2022

Author: Wang

Notes:

when fov is larger than 180° reaches negative plane -> [u,v,1]^T cannot be used
feature points rps with unit vectors
public fisheye dataset PALVIO
SVO, VINS Mono, ORBSLAM 3 support fisheye camera models
KLT sparse flow, RANSAC
Epipolar geometry to initialize, + Epnp to initialize depth of features

Panoramic SLAM:

PAL SLAM with panoramic annular cameras
many converts it to pinhole model => doesn't enable very wide FOV

System:

Shi Tomasi Corners
Unified spherical model for cameras with OmniCalib toolbox (Scaramuzza)
Epipolar -> Triang -> Epnp -> global BA
classic init procedure

Experiments:

LVI SAM dataset for public fisheye dataset (single camera...)
investigating different FOV => whole FOV offers best performances
Comparison with SVO2.0 and VINS-Mono
Future work: adding stereo and LiDAR