Predicting walking-to-work using street-level imagery and deep learning in seven Canadian cities

Abstract

New 'big data' streams such as street-level imagery are offering unprecedented possibilities for developing health-relevant data on the urban environment. Urban environmental features derived from street-level imagery have been used to assess pedestrian-friendly neighbourhood design and to predict active commuting, but few such studies have been conducted in Canada. Using 1.15 million Google Street View (GSV) images in seven Canadian cities, we applied image segmentation and object detection computer vision methods to extract data on persons, bicycles, buildings, sidewalks, open sky (without trees or buildings), and vegetation at postal codes. The associations between urban features and walk-to-work rates obtained from the Canadian Census were assessed. We also assessed how GSV-derived urban features perform in predicting walk-to-work rates relative to more widely used walkability measures. Results showed that features derived from street-level images are better able to predict the percent of people walking to work as their primary mode of transportation compared to data derived from traditional walkability metrics. Given the increasing coverage of street-level imagery around the world, there is considerable potential for machine learning and computer vision to help researchers study patterns of active transportation and other health-related behaviours and exposures.

Figure 1

Example of six horizontal Google Street View images for a given postal code location*. *The image segmentation algorithm used all images (images 1 to 6) while the object detection algorithm used every second image (images 1, 3, and 5) to avoid counting the car twice. Images extracted from Google Street View: © 2002 Google.

Figure 2

Pearson correlation coefficients for log-transformed walk-to-work rate and pixel coverage (IS) and counts (OD) of different features derived from GSV images within different buffer distances from postal codes*. *Walk-to-work rates are for postal codes with > 0% reported walk commuting. Horizontal dashed lines show > 0.45 and − 0.45 correlation thresholds used to identify variables for inclusion in subsequent regression analyses.

Figure 3

Linear relationships of GSV features within 1500 m from postal code with log-transformed walk-to-work rates.