FutureStructure

Using Crowdsourced Dashcam Videos to Map Roads for Robot Cars

A small San Francisco startup says it has a cheap, quick way to create high-resolution, detailed maps of millions of miles of roads: computer-vision software that can generate highly accurate maps from crowdsourced dashcam video.

by Carolyn Said, San Francisco Chronicle / July 20, 2017

(TNS) -- Digital maps will help autonomous cars navigate the world, pinpointing not just roads but every stop sign and traffic light. They’ll need to be constantly maintained as roads and conditions (think potholes) change. They’ll need centimeter-level accuracy, far more fine-grained than current maps in car navigation systems, which are accurate only within yards.

Creating high-resolution, detailed maps of millions of miles of roads is no small feat. Several well-funded players are hard at work on the challenge. Generally they drive around cars equipped with lidar, a laser form of radar, to capture 3-D images of streets and environs. That’s an expensive and time-consuming task.

Now a small San Francisco startup says it has a cheaper, quicker way to do it: computer-vision software that can generate highly accurate maps from crowdsourced dashcam video. It’s paying professional drivers — such as Uber, Lyft and truck drivers — to mount smartphones on their windshields to collect the video.

“We’re trying to make the best maps for all the companies working on self-driving,” said Andrew Kouri, 25, CEO and co-founder of Lvl5. (The company’s name is pronounced “Level 5,” which is the designation for cars that can drive themselves in all circumstances.)

“We can create 3-D maps from 2-D video, based on hundreds or thousands of trips down a given road, and get the same accuracy as lidar-created maps,” Kouri said, showing colorful 3-D point clouds on computer screens. “Our approach can scale everywhere in the world.”

Backed by $2 million in seed funding, Lvl5 is a recent graduate of the Y Combinator accelerator. Kouri and co-founder Erik Reed, 28, previously worked on Tesla’s Autopilot team, while co-founder George Tall, 26, was a computer vision engineer at iRobot, maker of the Roomba robot vacuum cleaner.

As the three explore cutting-edge computer science to enable a Jetsons-like future, they live and work together in one of the oldest houses on Russian Hill, the 1858 Feusier Octagon House, a San Francisco landmark and one of only two surviving octagon houses in the city. The company’s big-screen monitors, Ikea furniture and Costco rugs seem a tad out of place in the stately surroundings. The three founders, plus one staff member who doesn’t live there, cook together in the remodeled kitchen, mainly bulk items from Costco.

Kouri says Lvl5 has mapped some 500,000 miles of U.S. roads, about a quarter of the nation’s total, in just three months. About 2,500 professional drivers have installed its Payver app and agreed to send Lvl5 data from their daily drives. (The company provides a little suction cup for the windshield to position the phones correctly). In exchange, Lvl5 pays the drivers from 2 to 5 cents a mile, with higher rates for less-frequented areas.

By contrast, German mapping giant Here has been working on HD maps for a couple of years and said it plans to have covered about 311,000 miles by year end, including all major highways in North America and Western Europe. Its employees harvest data by driving some 400 mapping cars worldwide topped with rigs containing $70,000 lidar sensors and other equipment.

“They way they’re approaching this is the yesteryear way of mapping,” Kouri said.

But Here, which is owned by automakers Audi, BMW and Daimler, with significant backing from Intel, has deep pockets and extensive resources. Its parent companies soon will embed technology in the millions of cars they manufacture to enable crowdsourced anonymized data to flow to Here.

Max Altman, a partner at 9Point Ventures, one of Lvl5’s investors, said he has confidence in the company’s approach.

“Even before meeting the team, I thought, ‘Holy cow, this will affects the economics of how autonomous vehicles of every type will eventually hit the market,’” he said. Cheaper, scalable maps “make autonomous vehicles much more mass-market feasible.”

Still, he said, much of his confidence is predicated on the assumption that lidar will stay expensive. Meanwhile a bevy of players are working to make the sensors significantly cheaper.

Mapping expert Ugur Demiryurek, associate director of the University of Southern California’s Integrated Media Systems Center, who is not involved with the company, said Lvl5’s approach sounds intriguing, but he wonders if it truly will attain the accuracy of lidar-generated maps. “Lidar is critical in measuring all the depths and different geometric figures and shapes, for instance, differentiating between trees and traffic light poles,” he said. “It also is the standard tool to sense road elevation and curves.”

Kouri said Lvl5 is close to landing some deals with automakers, but can’t give specifics. His company has the ability to spring into action almost instantaneously, he said.

“Say Toyota came to us tomorrow and asked for maps of Japan,” Kouri said, speaking hypothetically. “We can turn on the whole country just by marketing to (professional) drivers there. Our scale happens overnight.”

©2017 the San Francisco Chronicle. Distributed by Tribune Content Agency, LLC