Automating City Accessibility Constraints Mapping Through AI-Assisted Scanning of Street View Imagery

Abstract

Urban environments often pose challenges for individuals with mobility impairments due to inadequate pedestrian infrastructure. In addition, the lack of accurate mapping of accessibility features limits the ability to monitor and address these constraints effectively. This paper introduces a framework for Automating City Accessibility Mapping using AI (ACAMAI), that is, provides an AI-assisted pipeline for the automated identification and geolocation of urban accessibility constraints using Google Street View (GSV) panoramas. The ACAMAI pipeline comprises two main stages: (i) training a YOLOv8 object detector to recognise accessibility-related features, such as curb ramps, missing ramps, obstacles and surface problems, in 2D sidewalk images; and (ii) scanning 360° GSV panoramas by extracting multiple perspective views to be analysed by the trained model. The model was trained on a combination of international (Project Sidewalk Dataset—PSD) and local (Porto Dataset—PTD) datasets, achieving high performance across classes, including 91% recall and 85% precision for curb ramps. In the panorama scanning stage, using a fine angular iterative step (2°) maximised the recall, reaching 90% for curb ramps and 93% for obstacles in a locally annotated dataset (GSV Panorama Porto Dataset—GSV-PPD). Although this improved detection coverage, it also led to a high number of redundant predictions, which contributed to a reduced overall precision. Finally, identified constraints are georeferenced and mapped onto OpenStreetMap (OSM), supporting scalable and inclusive urban planning.