Real-time Monocular Visual Odometry for Turbid and Dynamic Underwater Environments

Author: Ferrera

Year: 2020

Notes:

Simulated and real data
Artificial lightening mandatory
turbidity effects due to floating particles
front-end with optical flow and back end with bundle adjustment (classic...)

Feature tracking:

LKT to be robust to repetitive patterns
Shi and Tomasi algorithm to detect Harris corners
Benchmark of keypoint + descriptors (ORB, SURF...) and LKT with Harris
The search for matching descriptors is in 40*40 pixel zone
Harris corner is the only one to detect enough features per images
Epipolar check + RANSAC
final solution: Harris with LKT

VO:

pb of LKT: only available for succesive frames. Not robust to dynamic env => need of a retracking mechanism
Filtering of pts with forward/backward optical flow and essential matrix estimation in a RANSAC fation
Retracking of lost features by keeping a small window of the most recent frames
P3P in a RANSAC fation + BA for pose estimation between frames
KF voting with parallax and number of corespondence
Adaptive selection of KF included in the BA pb

Results:

SVO and LSD require high frame rate
LSD fails because it is based on strong gradient that are not frequent underwater
groundtruth with SFM software Colmap
BoW doesn't work for loop closure

Commentaire:

Système trés similaire à PAVO mais bien rédigé. Continuer de lire des paps sur la VO sous marine, un bon exemple de recherche applicative du SLAM