EG-ARSA: An Expert-Grounded Open Model for Visual Road Safety Auditing in Low-Resource Settings
Researchers propose Expert-Grounded Distillation (EGD), an AI framework that transfers institutional road safety expertise into a compact vision-language model for visual road safety auditing. The teacher model is calibrated against authoritative field audits, achieving Cohen's kappa = 0.74 before generating supervision. The student is an 8-billion-parameter vision-language model trained with Low-Rank Adaptation and a single leakage-free prompt. The work introduces BD-ARSA, an open expert-grounded Bangladeshi visual road safety audit dataset with 21,947 image-audit records and near-national coverage.
EG-ARSA is an expert-grounded open model for visual road safety auditing in low-resource settings. It uses Expert-Grounded Distillation to transfer institutional road safety expertise into an 8-billion-parameter vision-language model. The teacher model is calibrated against authoritative field audits, reaching Cohen's kappa = 0.74 before generating structured supervision. The framework also introduces BD-ARSA, the first open expert-grounded Bangladeshi visual road safety audit dataset containing 21,947 image-audit records with near-national coverage.
The key technical contribution is Expert-Grounded Distillation, which quantifies teacher agreement with expert risk assessments before allowing large-scale annotation. The student model is an 8-billion-parameter vision-language model trained with Low-Rank Adaptation and a single leakage-free prompt. The dataset BD-ARSA provides 21,947 image-audit records with near-national coverage, enabling scalable visual road safety auditing in low-resource settings.
This work addresses a critical gap in road safety auditing for low- and middle-income countries, where incomplete crash records and shortages of qualified auditors limit proactive safety measures. By providing an open model and dataset, it lowers barriers for institutions to adopt AI-assisted road safety auditing. The expert-grounded approach may serve as a template for other domains requiring institutional expertise transfer.
The open model and dataset can reduce the cost of large-scale road safety inspections and enable proactive auditing in regions with limited expert resources. This may create opportunities for public-private partnerships, consulting services, and technology transfer to government agencies focused on road safety.
Observable next signals include adoption of BD-ARSA by road safety agencies, further validation of the model on additional regions, and potential extension of Expert-Grounded Distillation to other low-resource auditing tasks. The release of the dataset and model may spur follow-up research on compact vision-language models for public safety applications.