EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

VI-BEVSEG: Vehicle-Infrastructure Fusion for Bird’s-Eye View Map Segmentation

Jingxiong Meng, Anirudh Ramasubramanian Iyer, Junfeng Zhao

IEEE Open Journal of Intelligent Transportation Systems · 2026 · 7 · 1573-1587

DOI: 10.1109/ojits.2026.3708074 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Open Journal of Intelligent Transportation Systems, 2026. Verfasst von Jingxiong Meng und 2 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With recent advances in infrastructure sensing, roadside cameras now provide wider-area coverage and vantage points that complement vehicle-mounted cameras for map segmentation. We present a camera-only, intermediate-fusion pipeline for map segmentation that jointly leverages infrastructure and vehicle visual streams. The system operates efficiently while maintaining strong accuracy, enabled by two key components: an Augmented Depth Estimation (ADE) block that refines monocular depth to produce more reliable BEV projections, and a Fast Cross-Attention (FC) block that efficiently fuses multi-view features to reduce computational cost without sacrificing accuracy. Our evaluation is conducted on a dataset primarily derived from V2X-Sim, supplemented with a custom ground-truth map-segmentation set aligned to the same scenes. Meanwhile, our work is further evaluated on a real-world dataset, V2X-Seq, to test out the model’s robustness and scalability. We follow standard segmentation protocols and report on IoU and computational time. Compared with a strong vehicle model, BEVFusion, our approach achieves notable gains in map-creation accuracy and extends the feasible mapping range, particularly in settings with distant structures and partial occlusions where infrastructure views offer complementary coverage. Our proposed VI-BEVSEG outperforms BEVFusion on V2X-sim by 2.68% in mIOU in the short-range case, and 3.25% in mIOU in the long-range case. Furthermore, when compared with the state-of-the-art V2I cooperative perception method UniV2X, our approach outperforms it by 4.55% in the short-range case and 7.43% in the long-range case. Source code link is: https://github.com/BELIV-ASU/VI-BEVSEG

Zitieren

APA

Meng, J., Iyer, A. R., & Zhao, J. (2026). VI-BEVSEG: Vehicle-Infrastructure Fusion for Bird’s-Eye View Map Segmentation. IEEE Open Journal of Intelligent Transportation Systems, 7, 1573-1587. https://doi.org/10.1109/ojits.2026.3708074

Vancouver

Meng J, Iyer AR, Zhao J. VI-BEVSEG: Vehicle-Infrastructure Fusion for Bird’s-Eye View Map Segmentation. IEEE Open Journal of Intelligent Transportation Systems. 2026;7:1573-1587. doi:10.1109/ojits.2026.3708074

BibTeX
@article{Meng2026,
  author = {Meng, Jingxiong and Iyer, Anirudh Ramasubramanian and Zhao, Junfeng},
  title = {VI-BEVSEG: Vehicle-Infrastructure Fusion for Bird’s-Eye View Map Segmentation},
  journal = {IEEE Open Journal of Intelligent Transportation Systems},
  year = {2026},
  volume = {7},
  pages = {1573--1587},
  doi = {10.1109/ojits.2026.3708074}
}
Kostenlos starten →