Soutenance de thèse - Ali Ahmadi - Doctorat en sciences géomatiques

Membres du jury

Titre

Towards City Scale Pedestrian Network Accessibility Assessment: An Artificial Intelligence and Multisensory Data Fusion Approach

Résumé

Approximately 8 million Canadians live with a disability, and about 39% report mobility-related limitations that restrict autonomous travel and social participation. For wheelchair users, discontinuous sidewalks, inadequate curb cuts, steep slopes, and poorly designed building entrances remain major obstacles. Yet information on these barriers is rarely available: accessibility is still assessed through manual audits that are fragmented, labor-intensive, and outdated, with no consensus on relevant factors. Single-modality data and manual characterization also prevent city-scale assessment. This thesis develops a scalable, automated approach for extracting quantitative accessibility parameters of pedestrian networks from multisensory 3D data, in four phases. First, a systematic scoping review (Arksey and O'Malley's methodology) identifies the environmental factors conditioning wheelchair mobility, isolates six dominant components (sidewalks, steps, curb cuts, ramps, crosswalks, streets), and reveals reliance on manual data collection and a lack of connectivity between building entrances and sidewalks. Second, Victoriaville 3D, an ultra-high-resolution dataset, is acquired with a Trimble MX50 mobile mapping system over 1,850 meters of urban streets, combining annotated LiDAR point clouds, 360-degree RGB-D imagery, and projected coordinate images into a new benchmark. Third, Y-Net, a dual-stream cross-fusion deep learning architecture, segments accessibility-critical elements with 96.92% overall accuracy and 70.15% mean IoU, delineating sidewalks, the primary wheelchair travel surface, at 84.88% IoU. Fourth, a modified Tangent Bug algorithm on a virtual triangular robot converts segmented point clouds into quantitative descriptors. Of the 13 parameters extracted, five characterize accessibility directly, namely running slope, cross-slope, path width, surface type, and step height, aligned with the Measure of Environmental Accessibility and ADA/IBC requirements; the others serve the robot's navigation logic. Path width is recovered with high accuracy (MAE = 0.027 m; NRMSE = 2.71%). Together, these phases form a pipeline from knowledge synthesis to deployment-ready accessibility information, supporting large-scale accessibility assessment and personalized assistive navigation for people with motor disabilities.