Street walkability inference with street view imagery through fine-tuning and prompting pre-trained large models

Abstract

Large-scale assessment of street walkability using street view imagery is essential for supporting urban spatial governance and achieving sustainable development goals. This study proposes a walkability inference framework based on pre-trained multimodal large models, which integrates domain-specific knowledge of the hierarchy of walking needs through fine-tuning and prompt engineering. This framework consists of three sequential steps: filtration of street view images, hierarchical diagnosis of street walkability, and treatment for walkable streets. First, LoRA fine-tuning is applied to align the model’s spatial perception standards with those of human experts, enabling independent assessment of accessibility, safety, comfort, and pleasurability, and facilitating the identification of spatial issues at each level. Second, prompt engineering is employed to embed the hierarchical logic of walking decision-making, allowing the model to determine priority improvement levels for the given street pedestrian environment and generate tailored spatial recommendations. To validate this approach, we conducted an empirical study, examining 17,937 street view images within the Inner Ring Road of Shanghai. The results confirm that our walkability inference model effectively performs logical inferencing based on the hierarchy of walking needs theory, addressing interpretability gaps and hierarchical deficiencies in street walkability assessment. Furthermore, the model generates context-aware, expert-level spatial recommendations in natural language, addressing the large-scale engineering requirements for improving pedestrian environments.