Research Summary: Can VLMs Reliably Assess Sidewalk Accessibility Attributes from Pedestrian-Level Imagery?

Research from arXiv explores the reliability of Vision-Language Models (VLMs) in assessing sidewalk accessibility attributes from pedestrian-level imagery, crucial for urban planning and compliance. The study, conducted using images from Seoul, South Korea, indicates that while conformal prediction methods can achieve desired coverage for VLM-based assessments, the informativeness of these assessments varies significantly across different attributes like effective width, longitudinal slope, cross slope, and pavement condition.

Why it matters

This research provides insights into the potential for leveraging advanced AI, specifically Vision-Language Models, for automated infrastructure assessment and regulatory compliance. It highlights the technical feasibility and current limitations of using such technology to improve urban accessibility, impacting strategic decisions related to resource allocation for infrastructure development and maintenance.

Key insights

• Sidewalk compliance, including attributes like effective width, longitudinal slope, cross slope, and pavement condition, is critical for urban accessibility, especially for wheelchair users and individuals with reduced mobility.
• The study investigated the ability of Vision-Language Models (VLMs) to reliably assess these sidewalk attributes using pedestrian-level imagery.
• It presents the initial application of sampling-based conformal prediction (CP) for VLM-based accessibility assessment, testing four VLMs on 514 sidewalk images with field-measured ground truth data.
• Conformal calibration successfully achieved the nominal 90% coverage across all models and attributes, demonstrating its technical validity.
• However, the informativeness of the calibrated regions differed, with effective width yielding the most informative results, suggesting variability in VLM efficacy for different assessment parameters.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.17882