EUVIMEDEUVIMED entdecken
Crossref

MSF-SLAM: Multicandidate LiDAR Registration and Quality-Aware Sensor Weighting for Wheeled-Robot SLAM

Shaofeng Han, Bohan Zang, Kang Li, Yurun Liu

IEEE Sensors Journal · 2026-11 · 26 (21) · 33468-33480

DOI: 10.1109/jsen.2026.3735974 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Sensors Journal, 2026-11. Verfasst von Shaofeng Han und 3 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Gefördert von: Fundamental Research Funds for the Central University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Wheeled inspection robots without satellite navigation must estimate their pose from onboard sensors whose reliability varies along the route. Three limitations of current practice are addressed. A single registration strategy cannot cover the alternating motion patterns of an inspection task; fixed fusion weights cannot react to short-term sensor degradation; and the planar prior of a wheeled chassis lacks a safe entry point into the estimator. A fusion system, multi-sensor fusion (MSF)-simultaneous localization and mapping (SLAM), is therefore proposed. Three complementary point-cloud registrations run in parallel at the front end, and a lightweight network predicts their fusion weights frame by frame. Training labels come from projecting a reference motion onto the weight simplex, without manual annotation. A quality score formed from sliding-window residuals drives a multiplicative weight update derived from a Kullback–Leibler regularized problem. Wheel odometry is admitted only after time and motion gating, and is projected onto the ground plane. Over seven Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) sequences, the fused front end lowered the mean relative translation error by 25.0% against a single generalized-iterative closest point (ICP) candidate and by 71.2% against a single tight-ICP candidate, and it outperformed both candidates on every sequence. Under a controlled extrinsic fault, adaptive weighting reduced the induced trajectory deviation by 24.6% while leaving the fault-free output unchanged. Trained on real-robot data alone, the network transfers without tuning to KITTI and to the multisensor multi-sensor and multi-scenario dataset for ground robots (M2DGR)-plus benchmark, where the full configuration attains the lowest mean relative pose error. On the onboard computer, the system runs above the 10 Hz light detection and ranging (LiDAR) rate.

Förderung

  • Fundamental Research Funds for the Central University

Zitieren

APA

Han, S., Zang, B., Li, K., & Liu, Y. (2026). MSF-SLAM: Multicandidate LiDAR Registration and Quality-Aware Sensor Weighting for Wheeled-Robot SLAM. IEEE Sensors Journal, 26(21), 33468-33480. https://doi.org/10.1109/jsen.2026.3735974

Vancouver

Han S, Zang B, Li K, Liu Y. MSF-SLAM: Multicandidate LiDAR Registration and Quality-Aware Sensor Weighting for Wheeled-Robot SLAM. IEEE Sensors Journal. 2026;26(21):33468-33480. doi:10.1109/jsen.2026.3735974

BibTeX
@article{Han2026,
  author = {Han, Shaofeng and Zang, Bohan and Li, Kang and Liu, Yurun},
  title = {MSF-SLAM: Multicandidate LiDAR Registration and Quality-Aware Sensor Weighting for Wheeled-Robot SLAM},
  journal = {IEEE Sensors Journal},
  year = {2026},
  volume = {26},
  number = {21},
  pages = {33468--33480},
  doi = {10.1109/jsen.2026.3735974}
}
Kostenlos starten →