Advancing Urban Accessibility through AI: Scalable Machine Learning Approaches to Pedestrian Path Network Mapping and Assessment

Abstract

Pedestrian paths are central to a healthy and accessible transportation network. A connected pedestrian path network map detailing the location, attributes, and connectivity of sidewalks, crossings, and curbs is essential to building an accessible transportation system. While automobile road networks have been extensively mapped, mapping the transportation network for the paths that serve pedestrians is inconsistent, incomplete, or missing. This fundamental lack of network information creates gaps in personal travel, accessible transportation analytics, and city planning. Typical mapping methods mainly rely on human surveyors’ collections and annotations. These methods are non-standardized, laborious, costly, unscalable, and difficult to keep current. In this dissertation, we address this problem by developing scalable machine-learning approaches for the generation of pedestrian path network maps and the assessment of paths and infrastructures in the pedestrian environment. We start by introducing a novel dataset that consists of aerial satellite imagery data, street map imagery data, and rasterized geographic information system (GIS) annotations for important classes in the pedestrian environment in multiple cities. The dataset can be used in many scene-understanding tasks for analyzing pedestrian environments. We also introduce an end-to-end artificial intelligence (AI) pipeline for inferring connected pedestrian path networks map using existing street network information and aerial satellite images. The pipeline uses a multi-input semantic segmentation network trained on our dataset to generate predictions for important classes in the pedestrian environment and uses these predictions together with existing street information to infer a connected pedestrian path network. Next, we develop an automated system for the mapping and assessment of sidewalk networks on mobile physical devices, commonly termed edge devices. Our system leverages advances in efficient neural networks, image sensing, Global Positioning System (GPS), and compact hardware to power sidewalk mapping on the edge. The physical system runs on a lightweight and low-power embedded device, facilitating deployment on any battery-driven mobility device, such as a powered wheelchair, stroller, or scooter for real-time on-device pedestrian paths mapping and assessment. Lastly, to aid model selection in practical applications, we propose segmentation evaluation metrics that decompose to explainable terms and are sensitive to over- and under-segmentation errors. This new approach confers additional desirable properties, including robustness to segmentation boundaries. We contextualized the application of our metrics in current model selection problems that arise in practice when attempting to match the context of use to region-based segmentation performance in supervised datasets. In summary, this dissertation introduces scalable approaches in response to the problem domain which requires the ability to perform fast, accurate mapping and assessments of the pedestrian paths network, through the production of open, shared data with limited reliance on human laborers.