Computer Vision for Scalable Walkability Assessment

Funding: NIH- and ASU seed grant-funded foundational research (Arizona State University) · NIH/NHLBI SBIR Phase I award to industry partner Urban Design 4 Health Inc., with ASU as a named collaborator ($275,660; 2023–2024)

The Ambiguity

The built environment — sidewalks, crosswalks, streetlights, curb cuts — shapes whether people can safely and comfortably be physically active in their own neighborhood, and physical inactivity is a major driver of obesity, heart disease, stroke, and Type II diabetes. Researchers already had a validated way to measure that environment in detail: the Microscale Audit of Pedestrian Streetscapes (MAPS), a tool for systematically rating walkability features. The problem was scale. MAPS audits require trained human raters walking or virtually reviewing every street segment, which makes them accurate but expensive, slow, and impractical to run across hundreds or thousands of neighborhoods.

The open question was whether a computer vision model could learn to see what a trained human rater sees accurately enough to trust the results. If so, are models generalizable enough to be usable at population scale?

Approach

The foundational research, conducted during my postdoctoral work, paired deep learning with Google Street View imagery to automate walkability audits. We built detection models (using an EfficientNetB5 deep learning architecture) for eight specific microscale features drawn directly from the MAPS Mini audit tool — things like sidewalks, sidewalk buffers, curb cuts, crosswalk markings, walk signals, bike symbols, and streetlights.

We used an iterative train-and-correct loop: train on one set of labeled images, evaluate performance against a separate validation set, and keep retraining until the model’s accuracy, precision and recall were satisfactory. We then used the finished models to audit the actual neighborhoods of 512 participants in the WalkIT Arizona trial, and checked whether what the model detected lined up with independent, established measures of neighborhood walkability.

Years later, that same underlying technology became the technical foundation for a new application: Arizona State University partnered with Urban Design 4 Health Inc. — a company with an existing, peer-reviewed health-prediction tool called N-PHAM — on an NIH SBIR Phase I award. The project, PED-PHAM, applied the computer vision models to a much larger scope: detecting pedestrian environment features from Google Street View imagery around 2,173 participant home locations across Baltimore, Phoenix, San Diego, and Seattle, drawn from two separate NIH-funded R01 studies, then feeding those features into Urban Design 4 Health’s existing physical-activity and BMI prediction models.

What We Found

The models performed well enough to be trustworthy, not just technically impressive: precision, recall, and overall accuracy for the eight detected features all exceeded 84%. More importantly, what the model detected wasn’t just internally consistent — it also tracked with real-world measures. Total detected microscale walkability was significantly associated with independently measured neighborhood walkability, and specific model-detected features (like sidewalks and sidewalk buffers) correlated positively with what residents themselves reported about their own neighborhoods.

That combination — strong classifier performance plus real-world validity across four different regions — signaled success. It meant the model wasn’t just automating a task, it was producing results a researcher could use in place of a costly human-rater audit or be unable to conduct at all.

What Changed

This work gave researchers a validated alternative to manual walkability audits — one capable of assessing hundreds or thousands of neighborhoods for population health surveillance or research, at a fraction of the cost and time of trained human raters.

The technology’s second life came through the PED-PHAM SBIR project. Its stated goal was to scale the approach to a national tool integrating AI-detected pedestrian environment features into physical activity and BMI prediction models for public planning agencies, consulting firms, developers, and healthcare and lending institutions — with a planned Phase II to add commercialization and air-pollution exposure modeling. Phase I was funded and completed (2023–2024); Phase II was never funded, so the work currently remains at the validated, Phase I research stage with work continuing via NIH funding, refining models for rural U.S. areas and expanding the number of model-detected pedestrian features. The lasting contribution is the validated technology and the research applications it has enabled.