EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Selective and Noise-Resistant Virtual Point Cloud Fusion for Long-Range 3D Object Detection

Chih-Ming Hsu, Yuan-Ting Fu, Bo-Yun Wang

IEEE Open Journal of Vehicular Technology · 2026 · 7 · 1399-1412

DOI: 10.1109/ojvt.2026.3689631 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Open Journal of Vehicular Technology, 2026. Verfasst von Chih-Ming Hsu und 2 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Virtual point clouds alleviate long-range sparsity in LiDAR-based 3D object detection by back-projecting image-guided depth into 3D space. However, depth uncertainties near object boundaries and occlusions introduce boundary-aligned artifacts, which can degrade detection if virtual points are fused without selection and can also increase runtime overhead due to excessive point density. This paper presents a selective and noise-resistant virtual point cloud fusion framework that explicitly suppresses boundary artifacts and improves cross-modal feature integration under practical efficiency constraints. The proposed method consists of three components. First, Dilated Noise-Resistant Convolution (DNrC) performs projection-domain context aggregation to mitigate boundary artifacts before cross-modal interaction. Second, Multiscale Cross Multi-Head Attention (MCMA) selectively injects multiscale image semantics into sparse 3D voxel features under calibration-consistent alignment. Third, a lightweight Cascade Attention Module (CasA) progressively refines predictions to improve localization for long-range and hard samples. Experiments on the KITTI validation set demonstrate consistent improvements over a strong virtual point baseline, with clear gains on challenging settings while maintaining practical inference speed.

Zitieren

APA

Hsu, C., Fu, Y., & Wang, B. (2026). Selective and Noise-Resistant Virtual Point Cloud Fusion for Long-Range 3D Object Detection. IEEE Open Journal of Vehicular Technology, 7, 1399-1412. https://doi.org/10.1109/ojvt.2026.3689631

Vancouver

Hsu C, Fu Y, Wang B. Selective and Noise-Resistant Virtual Point Cloud Fusion for Long-Range 3D Object Detection. IEEE Open Journal of Vehicular Technology. 2026;7:1399-1412. doi:10.1109/ojvt.2026.3689631

BibTeX
@article{Hsu2026,
  author = {Hsu, Chih-Ming and Fu, Yuan-Ting and Wang, Bo-Yun},
  title = {Selective and Noise-Resistant Virtual Point Cloud Fusion for Long-Range 3D Object Detection},
  journal = {IEEE Open Journal of Vehicular Technology},
  year = {2026},
  volume = {7},
  pages = {1399--1412},
  doi = {10.1109/ojvt.2026.3689631}
}
Kostenlos starten →