A study on the development of an automated motorcycle helmet wearing detection system via fine-tuning of the LLaVA-1.6 model
Journal of Data Forensics Research · 30.06.2025 · 2 (1) · 37-54
DOI: 10.12972/jdfr.2025.2.1.3 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Data Forensics Research, online erschienen am 30.06.2025. Verfasst von Taeyeon Kim und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Although existing object-detection studies can identify the 'presence' of a helmet, they often fail to contextually assess whether the helmet is actually being worn. To address this limitation, we developed a visual question answering (VQA) system based on LLaVA-1.6, a state-of-the-art vision-language model (VLM), to determine the helmet-wearing status of motorcyclists. A custom dataset of 1,555 image-question-answer pairs was constructed using public datasets, specifically designed to elicit clear "Yes” or ”No" responses to a fixed question. We fine-tuned the model using Quantized Low-Rank Adaptation (QLoRA), a parameter-efficient technique that specializes the model by only updating a small set of adapter weights. Experimental results showed that the proposed system achieved 95% accuracy, significantly outperforming the vanilla model, which reached only 88% accuracy and required complex heuristic post-processing. Importantly, the system notably reduced critical false-negative errors—cases in which wearers were misidentified as non-wearers— highlighting its reliability for real-world enforcement scenarios. In conclusion, this study demonstrates that task-specific optimization of VLMs can enhance the applicability of AI in traffic safety, improving both the efficiency and accuracy of automated enforcement systems.
Zitieren
APA
Kim, T., & Kim, K. (2025). A study on the development of an automated motorcycle helmet wearing detection system via fine-tuning of the LLaVA-1.6 model. Journal of Data Forensics Research, 2(1), 37-54. https://doi.org/10.12972/jdfr.2025.2.1.3
Vancouver
Kim T, Kim K. A study on the development of an automated motorcycle helmet wearing detection system via fine-tuning of the LLaVA-1.6 model. Journal of Data Forensics Research. 2025;2(1):37-54. doi:10.12972/jdfr.2025.2.1.3
BibTeX
@article{Kim2025,
author = {Kim, Taeyeon and Kim, Kyungjong},
title = {A study on the development of an automated motorcycle helmet wearing detection system via fine-tuning of the LLaVA-1.6 model},
journal = {Journal of Data Forensics Research},
year = {2025},
volume = {2},
number = {1},
pages = {37--54},
doi = {10.12972/jdfr.2025.2.1.3}
}
EUVIMED