Training Computers to See the Built Environment Related to Physical Activity: Detection of Microscale Walkability...

Area of Science:

• Urban Planning
• Computer Science
• Public Health

Background:

• Assessing neighborhood walkability is crucial for promoting physical activity but traditional manual audits are time-consuming and costly.
• Microscale streetscape features significantly influence pedestrian behavior and physical activity levels.
• Existing methods for streetscape feature detection are often resource-intensive, limiting large-scale application.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for automated detection of microscale streetscape features relevant to pedestrian physical activity.
• To leverage Google Street View (GSV) imagery for efficient and scalable streetscape audits, overcoming limitations of traditional methods.
• To evaluate the model's performance and its association with neighborhood walkability metrics.

Main Methods:

• Utilized the EfficientNETB5 architecture to train DL models for identifying eight key microscale features (sidewalks, buffers, curb cuts, crosswalks, signals, bike symbols, streetlights) using the Microscale Audit of Pedestrian Streetscapes Mini tool.
• Employed a train-correct loop with separate training and validation datasets to achieve high performance metrics.
• Applied the trained models to audit 512 participant neighborhoods in the WalkIT Arizona trial and explored correlations with macroscale walkability data.

Main Results:

• The DL models achieved high performance, with precision, recall, and overall accuracy exceeding 84% for all microscale features.
• A significant positive association was found between the total count of detected microscale features and overall macroscale neighborhood walkability (r = 0.30, p < 0.001).
• Model-detected sidewalks (r = 0.41, p < 0.001) and sidewalk buffers (r = 0.26, p < 0.001) showed significant positive correlations with self-reported data.

Conclusions:

• The developed computer vision approach provides a viable and efficient alternative to manual streetscape audits for trained human raters.
• This automated method enables large-scale audits of hundreds or thousands of neighborhoods, facilitating population surveillance and hypothesis testing in urban planning and public health research.
• The findings support the use of DL and GSV imagery for objective assessment of environmental factors influencing physical activity.