EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Semantic-Guided Depth Representation for Multi-Weather RGB–Depth Road-User Detection

Lija Akter, Mohammad Robel Miah, Ahmed Abdelmoamen Ahmed, Lening Wang

Future Transportation · 07.11.2026 · 6 (6) · 261

DOI: 10.3390/futuretransp6060261 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Future Transportation, online erschienen am 07.11.2026. Verfasst von Lija Akter und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Center for Infrastructure Transformation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Robust road-user detection in adverse weather conditions is essential for reliable autonomous driving, yet conventional RGB–Depth fusion does not explicitly emphasize the geometric regions associated with safety-critical traffic participants. This work proposes a semantic-guided RGB–Depth fusion framework that selectively modulates depth values within traffic-participant regions while preserving background geometry. A synchronized CARLA dataset comprising 6000 samples across ClearNoon, RainyNoon, FoggyNoon, and Night conditions was developed. The resulting four-channel representation is processed using an adapted YOLO26m-P2 detector, with controlled ablation experiments evaluating semantic-guided depth and high-resolution P2 detection. On the Town10HD test set, the framework achieves 94.39% precision and 95.99% mAP@50; semantic-guided depth improves mAP@50 and mAP@50–95 by 1.66 and 6.43 percentage points, respectively, over raw RGB–Depth under the same P2 architecture. Cross-town evaluation on unseen Town05 achieves 91.49% mAP@50 without retraining or fine-tuning. At a confidence threshold of 0.30, the system maintains 91.34% recall and detects 65 of 71 reference STOP events, while the detector operates at 48.80 FPS. The framework is evaluated under two distinct operating paradigms: an oracle reference condition utilizing simulator ground-truth semantic masks (achieving 91.13% mAP@50 and 81.67% mAP@50–95), and a realistic deployable condition utilizing on-board SegFormer-b0 predicted masks (achieving 71.52% mAP@50 and 49.07% mAP@50–95). Computational profiling on an NVIDIA RTX 5090 demonstrates a matched algorithmic perception-to-action execution latency of 34.85 ms (~28.7 FPS), within which the downstream rule-based safety computation accounts for only 0.030 ms (0.09%). These findings demonstrate the potential of semantic-guided RGB–Depth fusion for multi-weather road-user perception while highlighting the need for segmentation-robust fusion and real-world validation.

Förderung

  • National Center for Infrastructure Transformation

Zitieren

APA

Akter, L., Miah, M. R., Ahmed, A. A., & Wang, L. (2026). Semantic-Guided Depth Representation for Multi-Weather RGB–Depth Road-User Detection. Future Transportation, 6(6), 261. https://doi.org/10.3390/futuretransp6060261

Vancouver

Akter L, Miah MR, Ahmed AA, Wang L. Semantic-Guided Depth Representation for Multi-Weather RGB–Depth Road-User Detection. Future Transportation. 2026;6(6):261. doi:10.3390/futuretransp6060261

BibTeX
@article{Akter2026,
  author = {Akter, Lija and Miah, Mohammad Robel and Ahmed, Ahmed Abdelmoamen and Wang, Lening},
  title = {Semantic-Guided Depth Representation for Multi-Weather RGB–Depth Road-User Detection},
  journal = {Future Transportation},
  year = {2026},
  volume = {6},
  number = {6},
  pages = {261},
  doi = {10.3390/futuretransp6060261}
}
Kostenlos starten →