Dash­cams map routes fast

Startup taps driv­ers for video to chart course for ro­bot cars

San Francisco Chronicle - - BUSINESS REPORT - By Carolyn Said

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 li­dar­cre­ated 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.

“The 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 af­fect 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.”

Photos by Liz Hafalia / The Chron­i­cle

Lvl5 co-founders An­drew Kouri (left) and Erik Reed live and work at the Feusier Oc­tagon House in San Fran­cisco.

A real-time map shows driv­ers ac­cu­mu­lat­ing map data for Lvl5.

Photos by Liz Hafalia / The Chron­i­cle

Erik Reed (left), Ge­orge Tall and An­drew Kouri say Lvl5 has mapped about 500,000 miles of U.S. roads, about a quar­ter of the na­tion’s to­tal, in just three months, much faster than maps gen­er­ated from lidar. Reed (left) and Tall set up a dash­cam be­fore...

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.