Statewide intersection geometry extraction using geographic information system and deep learning model from road maps

Intersection/junction geometry, a crucial element in the Model Inventory of Roadway Elements (MIRE) Fundamental Data Elements (FDE), is required by the Federal Highway Administration (FHWA) to be comprehensively collected by all states by 2026. This element is not only vital for compliance with FHWA mandates but also significantly impacts traffic and demand management research. To assist the Nevada Department of Transportation (NDOT) in collecting this data for the entire state of Nevada, this article proposes an innovative framework, relying solely on the state's road network data. Our method involves extracting potential intersection points and their location information from the road network using a suite of geoprocessing tools in GIS, followed by acquiring corresponding road map images from network providers. These maps are then processed through the YOLOv5 model for intersection detection and classification, and the results are enhanced further by several postprocessing steps. Our approach, in stark contrast to traditional methods like manual collection or from vehicle-mounted sensors (camera or LiDAR), enables the rapid extraction of statewide data in just a few hours, significantly reducing the time required and overcoming the challenges associated with complex or large intersections. The precision of the YOLOv5 model has been validated, and a comparison with NDOT's existing partial state intersection records shows our method to be highly promising. It not only streamlines the collection process for MIRE FDE elements but also provides substantial assistance in updating existing state records.