Pioneering study uses traffic cameras and AI to predict future, promising to save lives and money

BY on

This special series focuses on important community issues, innovative solutions to societal challenges, and people and non-profit groups making an impact through technology.

Bellevue’s traffic cameras captured near miss events, tracking the speed and direction of vehicles, and by how many seconds they missed colliding. (City of Bellevue and Transoft Photos)

In an effort to prevent deaths and injuries caused by crashes between vehicles, bikes and pedestrians, the city of Bellevue, Wash., set out more than five years ago to foresee the future. The idea was to use machine learning to analyze thousands of hours of video collected by 360-degree traffic cameras already installed citywide to identify potential crash hot spots and tease out dangerous traffic patterns.

The city this week shared its first batch of reports from the groundbreaking program, which has drawn comparisons to “Minority Report,” the 2002 film in which Tom Cruise anticipates and preemptively stops crime.

“Bellevue is leading the way on this,” said Roger Millar, Washington Secretary of Transportation. “The analytics and AI going into it is doing a lot to predict and eliminate crashes.”

The project could aid Bellevue and other communities internationally that are aiming for “Vision Zero” — a strategy started in Sweden with the goal of eliminating deaths and serious injuries from traffic collisions.

The Bellevue Department of Transportation project examined collision close calls for vehicles, cyclists and walkers and calculated vehicle speeds. The study demonstrated a strong correlation between real-life crashes and the near misses detected from the deep neural network, validating the usefulness of the approach. It found overall traffic safety trends as well as identifying the most dangerous intersections.

Vancouver, B.C.-based Transoft Solutions, a transportation technology company, did the video analysis. The non-profit Together for Safer Roads paid for the study.

Among the findings:

  • While cyclists make up less than 1% of road users, they’re 10 times more likely to be involved in near miss collisions than motorists.
  • Intersections in residential areas had higher rates of speeding drivers and near misses compared to commercial areas.
  • Speeding motorcyclists were responsible for more crash close calls than anyone else on the road.
  • More than 10% of drivers were found speeding, and half were more than 11 miles per hour over the speed limit.
  • Two intersections of concern for high rates of near crashes between motorists and cyclists or pedestrians are in Bellevue’s downtown and near offices for tech companies including Microsoft, Amazon, Smartsheet and SAP Concur.

In the U.S., road crashes are the leading cause of death for people age 54 and younger. Nearly 40,000 people die annually in crashes and 4.4 million suffer injuries that require medical attention. In addition to the loss of human life, collisions rack up billions of dollars in public and private costs, including medical care, emergency services, property damage, lawsuits and lost workplace productivity.

A senior transportation engineer with the city of Bellevue at work in their transportation management center. (City of Bellevue Photo)

The traditional approach to road safety was documenting crashes over many years, waiting for problem sites to get noticed and performing labor-intensive field studies or consulting other research that might not be particularly applicable in search of solutions. The AI-enhanced video analysis creates a massive shortcut.

Historically, “it takes five-plus years for a pattern to emerge,” said Franz Loewenherz, principal transportation planner for the city of Bellevue.  “We’re able to compress five years into one week.”

Bellevue principal transportation planner Franz Loewenherz and project manager for the Video Analytics Towards Vision Zero Partnership.

The Bellevue study analyzed footage collected over 4,500 hours during one week in September last year, including observations of 8.25 million road users. The research focused on 40 intersections throughout the city, tallying crash close calls — which the transportation sector calls “conflicts,” with “critical conflicts” referring to the closest of close calls.

During that week, across those intersections, there were 20,000 critical conflicts, where vehicles, bikes or people missed smashing into each other by only two seconds.

In a city that earlier this year sparked privacy concerns over a program that shared reports of violators of the state’s COVID-driven “stay home” order, some might worry about the data collected by the cameras. Loewenherz said the analysis does not gather license plate information, use facial recognition or provide information to law enforcement.

Bellevue has already installed a new traffic light to provide protected left turns at one intersection red-flagged by the study. A followup analysis found that the change reduced critical conflicts by 60%. The city will be deciding on a new budget later this year and could include funding for more improvements prompted by the research.

“Results of the video analytics partnership work is great news for all who want to improve traffic safety,” said Bellevue Mayor Lynne Robinson by email. “If we can utilize technology and Bellevue’s cutting-edge traffic cameras to help predict where crashes will happen, we may be able to prevent them. Stopping incidents before they happen is the approach we need as we strive to reach our Vision Zero goal.”

AI-powered analysis of Bellevue’s the traffic camera videos has already led to the installation of a left turn signal at an intersection flagged for many near miss crashes. (City of Bellevue Photo)

The project kicked off five years ago with a partnership between Bellevue and Microsoft Research, whose engineers worked to develop software to analyze the traffic camera footage to categorize vehicles and pedestrians. In 2017, the University of Washington joined the project, and the team tapped the public to crowd-source some videos in order to help train the computer models. Despite making progress, Microsoft axed the project in 2018.

The infrastructure needed for this sort of monitoring is expensive, and few communities can match Bellevue for its widespread deployment of cameras, or its network for managing the giant volume of data produced. The cameras alone cost $3,000 to buy and install; Bellevue has cameras at 120 of its intersections and plans to cover all 200 intersections by next year. But governments including Oregon’s Department of Transportation and Washington County west of Portland are exploring the use of traffic cameras and analytics through temporary camera installations.

Bellevue’s research should provide information that’s transferable to other regions eager to improve road safety.

With videos capturing the moments leading up to collisions or near misses, “they’re able to see things that otherwise you wouldn’t have seen in accident records,” said Randy McCourt, president of the Institute of Transportation Engineers (ITE), an international professional organization. “There’s a lot to be revealed, and a lot of this is out on the cutting edge.”