BEVMatch: Uni-, Cross-, and Multi-Modal Keypoint Matching for Bird's-Eye View Perception
IEEE Robotics and Automation Letters · 2026-12 · 11 (12) · 13885-13892
DOI: 10.1109/lra.2026.3746099 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Robotics and Automation Letters, 2026-12. Verfasst von Lukas Beer und 3 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: dtec.bw—Digitalization and Technology Research Center of the Bundeswehr, European Union—NextGenerationEU, University of the Bundeswehr Munich.
Abstract (Originalsprache)
In this paper, we propose BEVMatch, a modality-flexible local correspondence representation in metric Bird's-Eye View (BEV) space for autonomous driving: camera and LiDAR observations are encoded into sparse landmarks with shared descriptors, enabling uni-, cross-, and multi-modal SE(2) matching with a single model. Using separate backbones for each modality, BEVMatch extracts features from each sensor, transforms them into a common BEV representation, and fuses them with cross-modal attention. Relative ego-pose estimates between two frames supervise the correspondences, so the model learns from large-scale driving data without manual keypoint annotations. Modality dropout during training encourages modality-consistent representations. Since every frame is encoded on its own, descriptors are not tied to one pairing but can be stored and matched against arbitrary later frames. We evaluate BEVMatch on nuScenes, Waymo and our in-house dataset across multi-camera-to-multi-camera, multi-camera-to-LiDAR, and LiDAR-to-LiDAR matching under large relative motions. We further demonstrate sparse, sub-cell-accurate landmarks whose keyframe odometry on KITTI yields trajectory-level errors comparable to reported BEV-based odometry. Code and trained models are available at https://github.com/UniBwTAS/BEVMatch.
Förderung
- dtec.bw—Digitalization and Technology Research Center of the Bundeswehr
- European Union—NextGenerationEU
- University of the Bundeswehr Munich
Zitieren
APA
Beer, L., Backhaus, A., Luettel, T., & Maehlisch, M. (2026). BEVMatch: Uni-, Cross-, and Multi-Modal Keypoint Matching for Bird's-Eye View Perception. IEEE Robotics and Automation Letters, 11(12), 13885-13892. https://doi.org/10.1109/lra.2026.3746099
Vancouver
Beer L, Backhaus A, Luettel T, Maehlisch M. BEVMatch: Uni-, Cross-, and Multi-Modal Keypoint Matching for Bird's-Eye View Perception. IEEE Robotics and Automation Letters. 2026;11(12):13885-13892. doi:10.1109/lra.2026.3746099
BibTeX
@article{Beer2026,
author = {Beer, Lukas and Backhaus, Anton and Luettel, Thorsten and Maehlisch, Mirko},
title = {BEVMatch: Uni-, Cross-, and Multi-Modal Keypoint Matching for Bird's-Eye View Perception},
journal = {IEEE Robotics and Automation Letters},
year = {2026},
volume = {11},
number = {12},
pages = {13885--13892},
doi = {10.1109/lra.2026.3746099}
}
EUVIMED