Marker-Constrained Pose-Graph Correction for Cross-Platform Georeferencing in GNSS-Denied Environments
2026-08-17 • Robotics
RoboticsMultiagent Systems
AI summaryⓘ
The authors developed a system that helps robots and drones know their precise location even when GPS is unavailable. They use special camouflaged markers placed in known positions that act like visual anchors to link different types of mapping data into one consistent map. Their tests showed these markers greatly reduce errors in positioning for both ground and flying robots. The method works quickly enough for real-time use and improves map accuracy without needing GPS signals.
GNSS-denied environmentsfiducial markersCholesteric Spherical Reflectors (CSRs)LiDAR-odometryRTAB-Mappose-graph optimizationgeoreferencingsimilarity alignmentUGVUAV
Authors
Marco Giberna, Jose Luis Sanchez Lopez, Holger Voos
Abstract
Autonomous operation in GNSS-denied environments requires heterogeneous mapping pipelines to maintain a consistent spatial reference. This paper presents a framework using camouflage-matched fiducial markers fabricated from Cholesteric Spherical Reflectors (CSRs) as pre-surveyed visual anchors. The anchors georeference both a lightweight LiDAR-odometry trajectory and a dense RTAB-Map reconstruction, allowing their outputs to be expressed in a common LUREF frame (geodetic coordinate reference system used in Luxembourg) without requiring GNSS measurements during operation. The method combines coarse similarity alignment with marker-constrained pose-graph optimization. We evaluate it using two handheld acquisition sessions with ground-level and elevated motion profiles emulating UGV and UAV operation. A single iMarker was relocated among six surveyed positions, with the first position revisited to quantify drift correction. Marker-anchor correction reduced revisit inconsistency by 97.9% and 99.1% for the UAV- and UGV-emulating sessions, respectively, and improved held-out anchor prediction compared with one-time alignment. Separately georeferenced dense reconstructions achieved a median cross-session nearest-neighbour distance of 58 cm without explicit cross-session registration. Marker processing operated in real time, while trajectory correction required less than 0.25 s per session. These results demonstrate a proof of concept for georeferencing lightweight odometry and dense reconstructions using visually unobtrusive, pre-surveyed anchors during GNSS-denied operation.