EUVIMEDEUVIMED entdecken
Crossref

A semantic-centered and stage-separated dynamic filtering strategy for RGB-D visual SLAM in dynamic indoor scenes

Jiupeng Chen, Xianjiang Li, Hongjun San, Fan Zhang, Wang Yang

Measurement Science and Technology · 17.04.2026 · 37 (16) · 166301

DOI: 10.1088/1361-6501/ae5d65 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 20.04.2026. Verfasst von Jiupeng Chen und 4 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Gefördert von: Yunnan Provincial Major Science and Technology Special Project, National Natural Science Foundation of China, Yunnan Province Basic Research Program Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Dynamic objects violate the static-world assumption in RGB-D visual simultaneous localization and mapping, often leading to ambiguous feature correspondences, pose drift, and degraded map quality. This paper presents CS-SLAM, a semantic-centered and stage-separated dynamic filtering framework that treats tracking and mapping differently according to their distinct spatiotemporal baselines and real-time requirements. In the tracking thread, a lightweight object detector deployed via TorchScript provides semantic priors, while sparse optical flow and epipolar-geometry constraints serve as geometric verification to detect slow-moving objects and suppress false positives. A confidence-stratified decision mechanism is further introduced to handle dynamic categories with different prior probabilities, enabling hierarchical filtering with semantics as the primary cue and geometry as auxiliary evidence. To mitigate the over-coverage of detection boxes and prevent excessive removal of static features, we propose a geometric anchor-guided bounding-box refinement that adaptively adjusts the filtering region when feature scarcity may cause tracking failure. In the mapping thread, multi-view depth-consistency checking with temporal verification is employed to suppress residual dynamic artifacts (e.g. ghosting) and improve spatiotemporal consistency, and an octree-based semantic mapping module is integrated for compact map representation. Experiments on TUM RGB-D dynamic sequences, the BOON dataset, and real-world indoor scenarios demonstrate consistent improvements in localization accuracy and map cleanliness over representative baselines under the same evaluation protocol. On five selected TUM sequences, CS-SLAM achieves an average absolute trajectory error root-mean-square error of 0.0200 m, with an average tracking time of 64.14 ms per frame under our implementation and hardware setup.

Förderung

  • Yunnan Provincial Major Science and Technology Special Project
  • National Natural Science Foundation of China
  • Yunnan Province Basic Research Program Project
  • Yunnan Province’s “Caiyun Postdoctoral” Innovation Project

Zitieren

APA

Chen, J., Li, X., San, H., Zhang, F., & Yang, W. (2026). A semantic-centered and stage-separated dynamic filtering strategy for RGB-D visual SLAM in dynamic indoor scenes. Measurement Science and Technology, 37(16), 166301. https://doi.org/10.1088/1361-6501/ae5d65

Vancouver

Chen J, Li X, San H, Zhang F, Yang W. A semantic-centered and stage-separated dynamic filtering strategy for RGB-D visual SLAM in dynamic indoor scenes. Measurement Science and Technology. 2026;37(16):166301. doi:10.1088/1361-6501/ae5d65

BibTeX
@article{Chen2026,
  author = {Chen, Jiupeng and Li, Xianjiang and San, Hongjun and Zhang, Fan and Yang, Wang},
  title = {A semantic-centered and stage-separated dynamic filtering strategy for RGB-D visual SLAM in dynamic indoor scenes},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {16},
  pages = {166301},
  doi = {10.1088/1361-6501/ae5d65}
}
Kostenlos starten →