Cluster-Based Loop Closing Detection for Underwater SLAM in Feature-Poor Regions

Author: Negre

Year: 2021

Notes:

Classical methods for loop closure fail in underwater environements
Loop correction necessary to limit the drift
idea: loop closure based on Keypoint clustering, (keypoint cluster can belong to several KF)
when matches are concentrated in a small area of the image, registration is not precise
No tracking stage in the SLAM system, it fails => use stereo odometer instead (egomotion taken from Stereoscan: Dense 3d reconstruction in real-time)
KF decision based on image overlap
Orb KP and clustering with DBSCAN, use stereo pairs to triangulate KP
Signature used is HALOC and it is based on SIFT descriptor (more efficient than BOW) and stored in a database
graph structure with both clusters centroids and odometry inputs (solved with g2o)
Loop detection by both proximity and cluster signature comparison
Uses the stereo 3D points of the cluster to perform a PnP Ransac 3D to 2D correspondence for every loop candidate, if there is enough inliers, a loop is validated (the number of inliers is used to make a weight in the graph)
Feature matching can be performed on several clusters from several past KF
Outperform odom only and ORB SLAM on underwater datasets with loops
KF to many KF loop mechanism enables many loops that are not detected with classic methods