Fully Automated Road Imagery Acquisition and Pavement Distress Mapping from Google Street View using Vision-Language Models: Proof of Concept

This paper presents a fully automated workflow for road surface imagery acquisition and zero-shot distress mapping from Google Street View (GSV). In a single run, the system samples a user-defined area of interest, retrieves forward-looking lane images, and queries a vision-language model (VLM) to detect pavement distress types and severities. The workflow outputs raw and labeled images, along with an interactive georeferenced map of pavement condition for each sampled location. We evaluate four GPT-based VLMs against three pavement engineering experts on 364 GSV images using standard reliability and accuracy metrics. The analysis quantifies VLM repeatability and agreement with a human consensus reference within this proof-of-concept dataset. Results show almost deterministic behavior, strong performance for the coarse distress-versus-no-distress decision, and high within-one-level severity agreement, although exact severity and full multi-class agreement remain below the human consensus. These findings support the feasibility of an automated, GSV-driven workflow for network-level pavement screening without task-specific training, while broader validation is still needed.