Road network classification based on street-level images and its machine learning embedding features

Abstract

This study introduces an approach for classifying road networks using geo-tagged street-level imagery, combining representation learning and clustering techniques. The method addresses scalability issues of traditional methods by gathering images from Google Street View. The process involves matching street-level images to road sections, extracting image features through a pretrained computer vision model, and applying clustering to categorize road sections. We apply this approach in Delft, Netherlands, classifying around 2,000 road sections with around 70 thousands images into six clusters, each representing distinct urban typologies. The clusters align with known road categories and reveal clear distinctions, such as residential, arterial, and motorway types. This method offers a scalable and adaptable solution, potentially improving urban planning, mobility studies, and automated vehicle navigation.