Driverless cars need particularly detailed mapping. What’s involved and how is it put together?
Onboard GPS has made getting lost hard for countless people in their cars. For driverless cars, though, instructions for navigation need to be more detailed than, say, “hard left, 200 metres”.
Companies including TomTom, NVIDIA and others are working to build out the required mapping detail for the approaching driverless future. High definition (HD) or highly autonomous driving (HAD) maps are maybe as close as we have come yet to 1:1 mapping of an environment.
“We believe that self-driving cars require a highly-detailed inventory of not just the road features, but things on the side of the road,” said Sharad Oberoi, director of IT & software engineering at Colorado Springs-headquartered Sanborn Maps.
“So not only a street, [but] all street signs, everything compiled in 3D, and our current collection sources don’t really provide that.”
Established in 1866, Sanborn originally made fire insurance maps and is the United States’ oldest continually operating map company. Today it is a “full-service geospatial provider” and is deploying its GIS infrastructure to develop the rich detail it believes is needed for autonomous vehicles. This includes representations of things such as lane markings, poles, footpaths, parking spaces, trees, signs and more.
Oberoi called what the company produces “reality-inspired models” for simulation and compares the approach of other car companies as more like a “video gaming mechanism” and less like accurate replicas.
“Sanborn has a large fleet of aircraft and a large portfolio of sensors. We primarily use Excel sensors,” explained John D. Russell, director of advanced technology at Sanborn.
“We use Phase 1 sensors. We use LIDAR sensors for the aircraft and therefore we were able to put together proprietary sensor packages that we use. And we use a combination of aircraft and mobile mapping systems.”
A combination of aircraft and mobile mapping systems are used, but the company estimates about 92 per cent of HD map data comes from the air.
Last year they undertook a project for an automotive company at Santa Clara (in Silicon Valley), where they had earlier completed a mapping project (in 2016) for three-inch oblique imagery.
The goals of the project were threefold, according to Oberoi: showcase the accuracy of their now-patented Roboroads HD mapping product, add a highly-detailed inventory of features onto the 3D mesh already collected, and finally ingest this all into a simulation program.
The Sunnyvale Heights district of Santa Clara was chosen as it is already used as a test area for companies making driverless cars, possesses a diverse mix of commercial, residential and municipal road networks, and features complex pedestrian areas and “urban route loop dynamics”.
Freeways are relatively simple for autonomous vehicles, but an environment like Sunnyvale requires “a whole different set of information”, according to Sanborn.
The 7-10 cm accuracy range is necessary, as it can be difficult to get clean GPS information in complex urban environments. Localisation is therefore trickier.
Russell, who heads a team of 16 developing HD maps, calls the degree of accuracy “on the edge” or just slightly outside of engineering-level mapping, but of a far greater accuracy than what “traditional navigation companies” offer, citing TomTom and HERE.
Mapping an area the size of Santa Clara is about a 90-day turnaround from collection to delivery. This has happened “in 2015, 2016, 2017 and our assets are collecting that area again next week,” says Russell when we speak. Regular updates of complex urban environments is necessary, and the company estimates there’s a 10-15 per cent change annually.
At the end of the project, measuring the Sunnyvale data’s precision involved the company’s expertise in aerial photogrammetry, said Oberoi.
“We did a whole bunch of control points on the ground and aerial triangulation reports and we actually proved that the accuracy is way better than the 7-10 cm accuracy required,” he said.
“It’s below 5 cm in most cases.”
For the moment, Sanborn is doing its HD mapping at a small scale. Auto manufacturers are not yet in a position to invest in large data sets and trials are only taking place in selected areas. Efforts in 2018 are being concentrated in California and Michigan.
As with any new industry, standards are still being hashed out. The biggest problem, according to Oberoi, is the level of detail clients will demand is not yet known.
“What I mean by that is we had one of our partners ask for the type of the tree on the side of the road: how does it matter if it’s a mango tree or some other tree, right?,” he said.
“But they didn’t know what they wanted, so they required so much detail that it becomes really costly to build these models. So right now we are hoping that between all the car companies they will come up with a standard specification.”