EG-ARSA: An Expert-Grounded Open Model for Visual Road Safety Auditing in Low-Resource Settings

Imagine you're working on improving road safety in a country where traffic accidents are common, but the data on these incidents is often incomplete and the resources for thorough inspections are scarce. Currently, road safety audits rely heavily on manual inspections and expert evaluations, which can be inconsistent and costly. This is what's called a lack of scalability, making it difficult to proactively address safety issues before they lead to accidents. The proposed solution is to leverage artificial intelligence to create a compact vision-language model that incorporates expert knowledge into a more efficient auditing process. By calibrating a teacher model against authoritative audits, the system ensures that the AI's assessments align closely with human experts before it generates structured supervision for a smaller student model. This approach allows for large-scale data annotation and effective training of the AI model, which is specifically designed for road safety tasks. The introduction of the Bangladesh Road Safety Audit dataset and the Expert-Grounded Road Safety Auditor model marks a significant step forward in making road safety audits more accessible and reliable in low-resource environments. Compared to previous methods, this framework not only improves the accuracy of risk assessments but also demonstrates that a smaller model can outperform larger counterparts, making it a practical solution for real-world applications.

The introduction of Expert-Grounded Distillation and a new dataset for road safety is a meaningful extension in the field.

The use of Cohen's kappa and comparisons with existing models provide solid evidence for the claims made.