Is It the Road's Fault?

• Author: Grabmeier, Jeff
• Position: The relation of serious traffic accidents to city roads

SERIOUS AUTO CRASHES in urban areas are more likely on city streets that look to drivers like highways, suggests a study from Ohio State University. Researchers applied machine learning techniques to analyze more than 240,000 images of road segments in Columbus taken from Google Street View. The goal was to see what the roads looked like to drivers and whether that was linked to serious and deadly crashes.

Findings showed street segments that were classified as "open roads"--those where the photos showed more visible sky, roadway, and signs--had 48% more crashes that caused injury or death than those classified as "open residential." The open road classification included almost all of the highway segments in Columbus (93%), but it also entailed more than half of the city's arterial road segments (59%)--the major high-capacity roads through the city.

"There are a lot of roads in urban areas that aren't highways but look like high-speed highways from the driver's point of view," says Jonathan Stiles, now at Florida Atlantic University, who led this work as a postdoctoral researcher in geography at Ohio State. "That's a problem because drivers behave as if those streets are highways, even though there may be a lot of pedestrians and human activity nearby."

The study was published in the journal Environment and Planning B: Urban Analytics and City Science. Coauthors are Harvey Miller, professor, and Yuchen Li, graduate student, both in the Department of Geography.

These dangerous "open roads" are familiar to anyone who has driven through urban areas, indicates Miller, who is director of the Center for Urban and Regional Analysis. They generally have multiple lanes of traffic, lots of road signs, and few trees, and are lined with strip malls, big-box stores, gas stations, and restaurants. 'To drivers, the road appears safe for driving at high speeds. We're combining a lot of complexity and human activity with the desire to move cars as quickly as possible. It is a dangerous combination."

After collecting the 241,179 street view images of road segments in Columbus, the researchers used a machine learning model to segment each photo into visual objects. The model was able to tell how much of each photo showed sky, roadway, signs, buildings, trees, and other objects. Sky was the most common element of the photos, covering 27.8% of the images, followed by road (24.6%), and trees (17.9%).

A computer "cluster analysis" identified four types of streets...