YawPR-Net: a yaw-robust network for large-scale point cloud based place recognition
Measurement Science and Technology · 01.05.2026 · 37 (18) · 186303
DOI: 10.1088/1361-6501/ae61dc ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 07.05.2026. Verfasst von Wen Hao und 2 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Shaanxi Provincial Department of Education Youth Innovation Team, China Scholarship Council.
Abstract (Originalsprache)
Point cloud-based place recognition plays an important role in large-scale robotic mapping and localization. However, existing methods often experience performance degradation when query point clouds undergo significant yaw rotations. To address this limitation, we propose a yaw-robust place recognition network for large-scale point clouds. The network adopts a pyramid feature extraction (PFE) architecture to learn multi-scale representations. Within each PFE module, a refined edge feature extraction unit is introduced to model local geometric relationships through a k -nearest neighbor graph, where an edge-wise attention mechanism adaptively emphasizes structurally informative relations. In parallel, a yaw-robust feature extraction unit constructs triangle-based geometric descriptors. Beyond conventional intrinsic attributes such as edge lengths and internal angles, orientation-aware components are incorporated by measuring angles between triangle edges and the gravity-aligned Z -axis. This design preserves yaw invariance while introducing vertical structural cues that help distinguish geometrically similar but differently oriented regions. The enhanced edge features and orientation-aware geometric features are fused hierarchically across scales, and a parameter-shared NetVLAD aggregation module is employed to generate compact global descriptors. Experimental results on benchmark datasets demonstrate that the proposed method achieves competitive performance and maintains stable recognition accuracy under large rotation variations.
Förderung
- National Natural Science Foundation of China
- Shaanxi Provincial Department of Education Youth Innovation Team
- China Scholarship Council
Zitieren
APA
Hao, W., Lv, Y., & Guo, L. (2026). YawPR-Net: a yaw-robust network for large-scale point cloud based place recognition. Measurement Science and Technology, 37(18), 186303. https://doi.org/10.1088/1361-6501/ae61dc
Vancouver
Hao W, Lv Y, Guo L. YawPR-Net: a yaw-robust network for large-scale point cloud based place recognition. Measurement Science and Technology. 2026;37(18):186303. doi:10.1088/1361-6501/ae61dc
BibTeX
@article{Hao2026,
author = {Hao, Wen and Lv, Yan and Guo, Li},
title = {YawPR-Net: a yaw-robust network for large-scale point cloud based place recognition},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {18},
pages = {186303},
doi = {10.1088/1361-6501/ae61dc}
}
EUVIMED