To Stop Distracted Driving, Researchers Monitor Drivers


Everyone knows that distracted driving is a problem, but it tends to fall in the “other people/not me” category of personal risk assessment among drivers. But when you consider that a staggering 80 percent of traffic accidents—and 17 percent of fatalities—are caused by distracted driving, according to the National Highway Traffic Safety Administration, that’s clearly flawed logic, by any measure. But while we’re confident that self-driving cars are on their way to save us from ourselves—however slowly—until they do arrive we have to deal with the fact that people are texting, tweeting, and just generally smartphoning at the wheel.

But a group of Canadian researchers think they can outwit those overconfident oversharers with the help of artificial intelligence. A team at the University of Waterloo’s Centre for Pattern Analysis and Machine Intelligence has developed software that can determine when drivers are texting or otherwise distracted—a potentially crucial step toward halting the habit.

“Driver distraction is a growing problem,” says program director Fakhri Karray, who studies electrical and computer engineering. Smartphones aren’t the only culprits: Today’s cars offer a bevy of infotainment features than can pull attention away from staying on the road. “If emerging electronic systems are not well-designed, they can become, and are becoming, new sources of distraction.”

Cars themselves could be less distracting, but automakers aren’t about to roll back high-tech features consumers like. Don’t expect people to suddenly develop self-discipline, either. The answer then, may be cars that can spot distraction in their drivers, whatever the cause.

University of Waterloo

That’s why Karray’s team created a prototype system that uses cameras—both Microsoft Kinect cameras and simple dashcams, mounted in a variety of locations on a simulated dashboard—to detect hand movements and algorithms to then grade them on how likely they are to put the driver in danger. That takes into account the act itself and its context, including the car’s speed, location, and driving conditions. Chatting on the phone while cruising on an empty highway may not be a huge problem. Reaching into the backseat while zipping down a busy boulevard probably is. If the system’s adequately alarmed, the car can give the driver an audio or visual warning. In the near-future, depending on how autonomous tech advances, the car could even take over control.

Car companies have already deployed distraction-tracking systems, mostly to ensure drivers remain attentive when their car is in semi-autonomous mode. Cadillac’s Super Cruise system, for example, tracks the human’s head position with an infrared camera. Other automakers are considering eye-tracking systems that know when a person’s actually watching the road, but the Waterloo team hopes to leapfrog past that solution.

  • Driven to Distraction

  • Aarian Marshall

    To End Distracted Driving, MIT Figures Out How People Really Drive

  • Jack Stewart

    Self-Driving Cars Are Confusing Drivers—And Spooking Insurers

  • Jack Stewart

    Self-Driving Cars Won’t Just Watch the World—They’ll Watch You

“The algorithm of the system we developed is powerful enough that it does not require the tracking of the human eyes or other facial landmarks,” says Karray. They created that algorithm with end-to-end deep learning, training the computer with a large number of images—hand positions, head placement—that involve known distracted-driving scenarios.

So how does this system know the difference between truly dangerous distraction and responsible glances at the radio or the passenger seat? Practice. “Unlike pattern recognition-based algorithms, deep neural networks learn from the huge number of samples presented to them to build their capabilities,” says Karray, who conducted the research with Waterloo’s Arief Koesdwiady, Chaojie Ou, and Safaa Bedawi. “The process is mostly autonomous, but it requires a large number of data, and significant computational capabilities. But deep learning has the lowest error rate, with the fewest false-positive and false-negative occurrences.”

Karray thinks while creating a standalone system based on his technology could be done in less than a year, integrating his program into production models would take longer, several years at least, as automakers figure out how to properly pester the distracted driver. But if and when the happens, just as many cars can now hit the brakes when they think a crash is likely, they’ll be able to respond when the driver goes mentally offline. More importantly, Karray says, this is a step toward giving cars a form of “self-awareness”—the equivalent, he says, to designing a cognitive artificial system. After all, if the car is paying attention to the world around it, it should probably keep tabs on what’s going on inside, too.