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

AID‐VLM: The Vision‐Language Model for Traffic Incident Understanding

Shuai Zhu, Yuchuan Du, Yishun Li, Chenglong Liu, Jun Wan, Difei Wu

IET Intelligent Transport Systems · 2026-01 · 20 (1)

DOI: 10.1049/itr2.70271 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IET Intelligent Transport Systems, online erschienen am 03.07.2026. Verfasst von Shuai Zhu und 5 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Key Research and Development Program of China, National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Traffic incidents threaten transportation safety and efficiency, yet current Automatic Incident Detection (AID) systems relying on handcrafted rules lack adaptability. To address this, we introduce AID‐VLM, a domain‐specific vision language model for traffic incident understanding. Our approach features three innovations: a self‐instruction data generation framework that integrates traffic domain knowledge to automatically create high‐quality, scenario‐diverse Q‐A pairs; Q‐LoRA for efficient fine‐tuning that drastically cuts computational costs by quantising model weights to 4‐bit precision; and O‐LoRA for incremental learning, enabling task transfer while retaining prior capabilities. AID‐VLM excels in detection, description, and deduction tasks, achieving a 75.3% average macro‐F1 score for incident detection. In preference evaluations, it outperforms baseline models with a 73.6% win‐or‐draw rate for descriptive narration and a 67.8% rate for causal deduction, generating more coherent, detailed, and causally grounded textual outputs. This work establishes a scalable framework for integrating interpretable VLMs into practical AID systems.

Förderung

  • National Key Research and Development Program of China
  • National Natural Science Foundation of China

Zitieren

APA

Zhu, S., Du, Y., Li, Y., Liu, C., Wan, J., & Wu, D. (2026). AID‐VLM: The Vision‐Language Model for Traffic Incident Understanding. IET Intelligent Transport Systems, 20(1). https://doi.org/10.1049/itr2.70271

Vancouver

Zhu S, Du Y, Li Y, Liu C, Wan J, Wu D. AID‐VLM: The Vision‐Language Model for Traffic Incident Understanding. IET Intelligent Transport Systems. 2026;20(1). doi:10.1049/itr2.70271

BibTeX
@article{Zhu2026,
  author = {Zhu, Shuai and Du, Yuchuan and Li, Yishun and Liu, Chenglong and Wan, Jun and Wu, Difei},
  title = {AID‐VLM: The Vision‐Language Model for Traffic Incident Understanding},
  journal = {IET Intelligent Transport Systems},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/itr2.70271}
}
Kostenlos starten →