Publication
Vision Without Surveillance: Using Zero-Shot CLIP for Scalable Built Environment Feature Extraction
| Summary: | Municipalities need current, granular built-environment data to support effective planning. However, current collection methods are expensive, challenging to scale, and present privacy concerns. To address this gap, the authors collaborated with the U.S. Department of Transportation (USDOT) through a Small Business Innovation Research (SBIR) initiative. We developed a scalable, privacy-preserving vision-language model (VLM) pipeline that leverages zero-shot CLIP to extract 150+ interpretable urban design features from street-level imagery. Applied to Philadelphia’s bus network, the pipeline processed over two million images and generated block-level predictions for more than 8,000 blocks. This approach enables scalable, evidence-based urban planning across cities. |
|---|---|
| Main Authors: | Rajshahi, Malikka |
| Other Authors: | Alfonzo, Mariela; Ashby, Corey; Taylor, Evan; Feuer, Benjamin |
| Subject: | Urban Planning AI for cities Vision-Language Models Planning support systems Urban metrics Scalable urban feature extraction |
| Year: | 2026 |
| Country: | Portugal |
| Document type: | article |
| Access type: | unknown |
| Associated institution: | Street Art & Urban Creativity Scientific Journal |
| Language: | English |
| Origin: | Street Art & Urban Creativity Scientific Journal |
| Summary: | Municipalities need current, granular built-environment data to support effective planning. However, current collection methods are expensive, challenging to scale, and present privacy concerns. To address this gap, the authors collaborated with the U.S. Department of Transportation (USDOT) through a Small Business Innovation Research (SBIR) initiative. We developed a scalable, privacy-preserving vision-language model (VLM) pipeline that leverages zero-shot CLIP to extract 150+ interpretable urban design features from street-level imagery. Applied to Philadelphia’s bus network, the pipeline processed over two million images and generated block-level predictions for more than 8,000 blocks. This approach enables scalable, evidence-based urban planning across cities. |
|---|
Funded activities
No funded projects found.