Aplicação de IA multimodal na inspeção da acessibilidade de calçadas: desempenho do GPT-4o na verificação dos parâmetros da NBR 9050

Abstract

This research investigated the feasibility of using computer vision and LLMbased processing to audit sidewalk accessibility in accordance with the NBR 9050:2020 standard, using a comparative validation methodology comparing on-site and computer-vision-based surveys. For data collection, 46 representative images of sidewalks in Juiz de Fora with varying characteristics were chosen. The development of the tests resulted in the “Sidewalk Analyzer” tool that combines GPT-4.1 processing with monocular geometry-extraction models. The research found that the isolated use of LLMs was insufficient for a fully reliable technical audit, but combining them with auxiliary models for geometric extraction showed significant potential to support public inspection. In addition to verifying the accuracy of the tested model, the research proposes a replicable methodological architecture that can evolve with technological advances.