Eyesight Technologies, a startup developing in-car driver and occupancy monitoring systems, today announced that its platform can now detect when drivers wearing face masks become distracted or drowsy. The upgrade comes as the U.S. Centers for Disease Control and Prevention recommends that people wear masks in public to prevent the spread of COVID-19, and as workers deemed essential by local governments, like truck drivers and bus operators, contend with longer-than-usual work hours.
In early March, the Federal Motor Carrier Safety Administration loosened restrictions on drivers hauling emergency loads such as medical and sanitary supplies, lifting an 11-hour-a-day cap on drive time. While drivers are still entitled a minimum of 10 hours off, Eyesight CEO David Tolub insinuates that the rule change could lead to more fatigued drivers on the road, which is particularly problematic considering it takes hundreds of feet to stop a low-speed 40-ton tractor-trailer.
“We are living in unprecedented times,” said Tolub. “Without a concrete end date to the current situation, wearing medical masks may be a reality for the foreseeable future. Eyesight … is forging ahead and adapting to provide a reliable solution to help guarantee safety even under less than ideal circumstances.”
Eyesight’s solution is a family of AI algorithms trained to monitor drivers through masks in addition to sunglasses, protective glasses, and visors. Both its Driver Sense product and Fleet Sense aftermarket system tap a combination of IR sensors and cameras for detection, enabling them to track drivers and their head position, eye openness, pupil dilation, blink rate, gaze direction, and more in all lighting conditions.
“Traditionally, [our systems] would … track certain factors on the mouth and face, but we have actively trained our AI and computer vision on use cases with a face mask to be able to perform even with these occluded,” an Eyesight spokesperson told VentureBeat via email. “We had to collect large amounts of data of drivers wearing various face masks in different driving scenarios. This is something that we had a headstart on while building our system for the Asian market, where masks have always been more prevalent.”
AI systems like those underpinning Eyesight’s products generally haven’t been developed with masks in mind, posing a challenge as the novel coronavirus pandemic rages on. For instance, Apple’s Face ID and the Google Pixel 4’s facial recognition feature recognize people if they’ve shaved their beard or put on sunglasses, but they’re rendered virtually useless by homemade cloth or medical masks.
In response, companies like Hanwang have turned to synthesis to generate data sets of masked faces, which have been used to train AI models to track and identify people in masks with surprising accuracy. In a case in point, researchers from Wuhan University released the Real World Face Recognition corpus — one of the largest data sets of its kind, ostensibly — and used it to train AI to achieve state-of-the-art performance, correctly recognizing people 95% of the time.
“More people covering their faces won’t meaningfully thwart face recognition technology,” wrote Northeastern University professor Woodrow Hartzog in an op-ed last week. “Companies are already creating workarounds that make educated guesses about what masked faces look like.”