AI Captures Subjective Walkability Perception with Multimodal Deep Learning.

Traditional studies on urban walkability often aggregate diverse individual judgments into single scores, assuming a uniform perception. This approach overlooks the significant variability in how people perceive walkability, which is influenced by personal characteristics, experiences, and preferences. Additionally, many studies rely on vehicle-mounted street-view imagery, which doesn't accurately reflect a pedestrian's visual experience. Researchers have addressed these limitations by creating a new dataset comprising 29,870 walkability ratings from 1,196 respondents, linked with sidewalk-view imagery across various Australian environments and individual rater attributes. They also propose the first user-conditioned multimodal deep learning framework for walkability perception, which fuses visual features with respondent-level representations. A comparative study revealed that sidewalk-view images consistently receive significantly higher walkability ratings than matched street-view images, underscoring the importance of imagery source. The user-conditioned model substantially improved rank agreement with observed ratings by 65% over an image-only baseline, demonstrating that incorporating "who" is evaluating an environment is as crucial as "what" is being evaluated. These findings advocate for a shift from generalized, observer-independent walkability scores to models that account for diverse user perceptions, enabling more inclusive and accurate assessments of pedestrian environments.