Deep Global Point Cloud Registration for Long-Term Robot Localization
Máximo, M.; Heredia-Aguado, E.; González-Amorós, E.; Martínez-Peral, F. J.; Payá, L.; Valiente, D. and Ballesta, M.
ICINCO 2026. 23rd International Conference on Informatics in Control, Automation and Robotics (Angers (France), 26-28 October 2026)  (2026)
Ed. ScitePress  ISBN:978-989-758-870-9  ISSN:2184-2809  DOI:10.5220/0015203000004979  BIBTEX:@conference{icinco26, author={Míriam Máximo and Enrique Heredia{-}Aguado and Esther González{-}Amorós and Francisco José Martínez{-}Peral and Luis Payá and David Valiente and Mónica Ballesta}, title={Deep Global Point Cloud Registration for Long-Term Robot Localization}, booktitle={Proceedings of the 23rd International Conference on Informatics in Control, Automation and Robotics - Volume 1: ICINCO}, year={2026}, pages={390-397}, publisher={SciTePress}, organization={INSTICC}, doi={10.5220/0015203000004979}, isbn={978-989-758-870-9}, issn={2184-2809}, }  - Volume 1, pages 390-397

Resumen:



Accurate localization and orientation are fundamental prerequisites for the autonomous navigation of mobile robots and vehicles. To this end, this paper evaluates a method capable of estimating a robot's pose with six degrees of freedom (6DoF) by comparing a real time LiDAR scan with a previously acquired scan of the same environment. Our approach relies on establishing local feature correspondences between the respective point clouds. These local features are extracted via a deep learning architecture originally trained to generate a global descriptor for the entire scan. This study analyses the performance of the model's intermediate layers when repurposed for this task. The analysis is conducted using standard state of the art datasets, supplemented by a proprietary collection of captured scans.