CECI and NTU Deploy AI and Multimodal Models to Evaluate Liugong Green Corridor Pedestrian Transformation

Urban public space enhancement is shifting from traditional facility provision toward human-centered environments. CECI Engineering Consultants, in collaboration with National Taiwan University and other researchers, established an AI-driven landscape and human-centered evaluation framework to quantitatively assess the Liugong Green Corridor project in Taipei. The study integrates Google Street View imagery, the LaDeco semantic segmentation tool, and the open-source multimodal large language model Gemma 3 using three-tiered prompt engineering with chain-of-thought guidance to convert subjective environmental perceptions into standardized metrics. Results show that spatial safety increased from 2.85 to 3.50, pedestrian network connectivity improved from 2.62 to 3.06, and universal accessibility rose from 2.54 to 2.97. Concurrently, LaDeco analysis revealed minimal changes in green view ratios, proving that pedestrian improvements stemmed from circulation restructuring and conflict mitigation rather than merely adding vegetation.

Key takeaways

• The framework uses Google Street View and the Gemma 3 multimodal model with prompt engineering across 18 indicators in six dimensions including safety, walkability, and comfort.
• Prompts incorporating expert chain-of-thought improved evaluation consistency in objective attributes, though abstract psychological metrics still faced cultural context limitations.
• In the Liugong Green Corridor evaluation, spatial safety improved by 0.65, universal accessibility by 0.43, and connectivity by 0.45, confirming tangible pedestrian gains.
• LaDeco visual segmentation revealed little change in vegetation proportions, indicating that improvements arose from space reorganization rather than simple greening.