In the U.S., car crashes kill roughly 40,000 people a year. That’s a loss of life equivalent to four to five 737 airplanes crashing every week.
“In no other setting would this level of fatality go unnoticed,” said Franz Loewenherz, principal transportation planner for the city of Bellevue, Wash.
What makes it even worse is “a lot of these are preventable crashes,” he said. “If we have the data, we can make smart choices about how to deal with this.”
So Loewenherz and the city have teamed up with Microsoft and the University of Washington to form the Video Analytics Towards Vision Zero Partnership. Their plan is to generate that data and use it to improve roadways and prevent collisions between cars, trucks, bikes and pedestrians. Their ultimate goal is eliminating traffic deaths and serious injuries, a target referred to as “vision zero.”
They’re going to need the public’s help to get there.
On Thursday morning, the group is launching a crowd-sourcing project that asks people to help them analyze pre-recorded traffic videos. Volunteers will label cars, bikes and people in videos and identify the trajectory of each. That information will be used to develop deep learning algorithms so that computers will be able to recognize the different objects and flag collisions and near misses.
By screening countless hours of traffic videos — a job too slow-going and laborious for people — the algorithm will eventually be able to pinpoint the most dangerous streets and intersections, and even give details on which days and hours the location is riskiest.
The research is a tremendous leap in the arena of transportation safety.
“Currently our approach is reactive,” said Loewenherz, who is also the project manager for the partnership. “And unfortunately we wait for people to get injured before having an indicator of where a problem is — and a lot of people have to be injured before a pattern emerges.”
The new approach is predictive, relying not only on actual crashes, but looking for near misses where someone brakes suddenly or takes evasive action.
“Computer vision techniques applied to video feeds from traffic cameras have a huge potential of improving traffic flow and reducing traffic crashes and fatalities,” said Victor Bahl, director of Microsoft’s Mobility and Networking Research, in a prepared statement.
“We are working diligently on this because we truly believe the societal impact will be significant,” Bahl said.
Microsoft and Bellevue initiated the collaboration more than two years ago. The technology company is contributing the time of roughly five engineers, who are working on software for analyzing the videos and writing algorithms to look for collisions and close calls. Bellevue is a local leader in vision zero efforts, and has worked for many years on a pedestrian and bicycle safety initiative.
The UW’s professor Yinhai Wang and his colleagues are also part of the effort. Wang is the founding director of the Smart Transportation Applications and Research Laboratory at the UW and director for Pacific Northwest Transportation Consortium. The UW scientists will work with the Microsoft engineers to correctly identify the near-miss events.
Other cities and organizations are supporting the effort. In the U.S., the cities of Seattle, Redmond, New York, Los Angeles, San Francisco and Gainesville, Florida, are partners on the project, as well as Snohomish and King counties in Washington. Canadian cities including Calgary and Vancouver, B.C. are onboard.
Participating nonprofits are the Institute of Transportation Engineers, Intelligent Transportation Society of America, Vision Zero Network, Cascade Bicycle Club and People for Bikes. University of British Columbia and McGill University are also playing a role.
Bellevue collects videos from 300 traffic cameras and other cities are also contributing their footage for the crowd-sourcing analysis. The videos are too low resolution to identify the people or read the vehicle license plates.
The project leaders caution that it may take volunteers five minutes or more to master the tools used to label the videos, and people are invited to analyze as few or many as they’d like.
Loewenherz didn’t know how long the crowd-sourcing project would run. It depends on how quickly the algorithm learns to correctly identify the images in the videos. The software currently recognizes vehicles about 95 percent of the time, but it accurately labels bikes and pedestrians 60-70 percent of the time. The software needs to reach 95 percent accuracy or higher.
The video analysis data would give transportation experts better information leading to a solution that targets the cause of the problem, Loewenherz said, saving lives and money. Cities could change the design of an intersection, install a roundabout, adjust traffic signals, boost police enforcement or run an education campaign to increase safety.
“Right now you rely on crash reports,” he said. “They don’t tend to give you enough data to make an informed decision.”