High-precision maps for self-driving cars

If cars are to drive themselves, they need to be able to determine their position exactly. High-precision maps that are updated and made available via the internet are being created to help them do this. Self-driving cars are very much in the public eye at present – while they may still be in the realm of the future, the future usually arrives quicker than expected. The next generation Mercedes-Benz E-Class, for example, which is launching in 2016, will feature additional functionality that enables semi-autonomous driving. If the driver asks it to, the E-Class will – at speeds of up to 200 km/h – follow the vehicle ahead, even around gentle bends in the road, detecting speed limits along the way and making sure they are not exceeded. Fully autonomous driving can be seen as an additional assistance function that in many situations will bring welcome relief for the driver – on routine journeys, for example, or in traffic jams. But to some extent it is also the culmination of all previous safety innovations targeted at accident-free driving, a vision that Mercedes-Benz has been systematically pursuing for a number of decades.

NAVIGATION SYSTEMS ARE NOT ENOUGH

Cars will only be able to drive themselves if they have access to high-precision maps. The digital material contained in today’s navigation systems is not enough. To be able to drive itself safely, a car needs to know its position on the road down to the centimetre. When turning, for example, the car cannot approximate the point at which the steering wheel should be turned, an area in which human drivers are experts at making adjustments. Digital driving instructions need to be ultra-precise. In addition to the map data, the self-driving car’s various sensors will, of course, also provide it with important information about its environment.

The Bertha Benz drive in the S 500 INTELLIGENT DRIVE research car in 2013 represented a major step towards the future. The luxury class saloon demonstrated that autonomous driving is possible both on cross-country journeys and in city traffic (see separate info box). Extensive preparations were made for the drive. This included creating the maps, as there was no existing material with the required precision. “A high-precision digital map was put together especially for the trip,” explains Martin Haueis, who heads up the Vehicle Tracking and Data Management department of Daimler’s corporate research and Advanced Development unit.

STEREO CAMERAS GATHER IMAGE DATA

The company worked with the Karlsruhe Institute of Technology (KIT) to create the map. “A stereo camera scanned the roads along the route, gathering image data. A high-precision GPS overlaid this with positional data, and the two together produced a highly detailed representation of the roads. It’s more like a 3D model of the world – we still call it a map, but it no longer has much in common with a traditional map. To improve the accuracy, we drove the route several times in order to give the adaptive system the necessary depth of data.”

Without the map, the Bertha Benz drive would never have succeeded. However, it became clear to the experts that the only practical way to guarantee that maps are accurate and up to date at all times is for them to be created on an ongoing and dynamic basis without any manual post-processing. “The real breakthrough will come when completely normal cars being used for everyday driving are gathering information about the roads,” says Haueis. “After all, they are continuously out and about in the places where mapping vehicles would otherwise have to be driven.” The disadvantage is that the data they record will not necessarily be as up to date as it needs to be because the cars might not drive on some roads again for another six months, for example – which isn’t enough for autonomous driving.

The experts set about improving the accuracy of the maps with the aid of self-learning systems. The initial aim was to have a vehicle create its own high-precision map. Test vehicles were fitted out with the necessary sensors and computer equipment. Drive by drive, they gathered the data from which an ultra-precise representation of the road and its immediate surroundings was put together.

VEHICLES GATHER MAPPING DATA

Stereo cameras build up a complete 3D image of the car’s periphery and identify features that will help in orientation – corners of buildings, for example, lampposts and road signs. Lane markings, directional arrows, pedestrian crossings, stop lines and kerbs are also detected. Together with the relevant positional data from the GPS a detailed representation of the route is created. This provides the basis for the autonomous drive and is always open for updates. The next time the car drives along that road, the map will automatically be improved.

“Valid base data is created the first time that a route is driven. If, say, two more trips are taken along the same route, the quality of the map will be brought up to a very high level,” explains Christoph Keller, who also works on vehicle tracking in Daimler’s advance development unit. Another advantage is that the map will grow as time goes on – whenever the car drives along a road it has never been on before it will create the relevant map data.

What sounds relatively straightforward actually poses numerous technical challenges. “After all, you won’t just get a good map simply because you have a lot of information,” says Keller. The data has to be processed in the right way in order to produce the level of detail that’s required.” Individual algorithms, executed by computer software, produced the desired result: a car programmed to learn by itself is now able to create its own high-precision digital map. This is important for autonomous driving but also for optimising routine trips from within the data-gathering car, which is seen as an additional convenience feature.

ALL FOR ONE, ONE FOR ALL

The next stages of development are already mapped out. The only practical way to guarantee efficiency and accuracy at all times is for vehicles to collaborate – i.e. to not create the maps individually but to do so centrally, using a multitude of pooled data that will then be processed and made available to all participating vehicles. “The vehicles will send the route data via mobile networks to a central computer, known as a backend because of how it works behind the scenes. This backend compiles all this data into a digital map that then finds its way back to the vehicles via the internet,” explains Haueis.

There are many advantages to this method. The maps used by the self-driving cars will have the highest level of detail and will always be up to date. And key resources such as computing power and learning algorithms will be centralised rather than having to be accommodated in the vehicles themselves. “For this solution too, the art lies in creating the necessary programming codes – as we’ve said, simply having a lot of data does not guarantee a high-quality result. What matters is how it’s processed.” The developers have not quite reached the point where they are satisfied with the backend coding. But they are close: The future does, after all, usually arrive faster than expected.

The Bertha Benz drive in 2013

The 100km drive from Mannheim to Pforzheim followed the same route taken by automotive pioneer Bertha Benz, who in 1888 made the first long-distance journey by motor car.

In mid-2013 the Mercedes-Benz S 500 INTELLIGENT DRIVE research car drove itself across country and in towns. On the busy roads of the 21st century the self-driving S-Class was able to handle highly complex situations – with traffic lights, roundabouts, pedestrians, cyclists and trams.

This game-changing success was unusual because it did not require expensive specialist technology to be developed, but was brought to fruition using equipment that is nearly production ready and is similar to what is already widely available in current Mercedes-Benz models.

The project represents a milestone on the journey from the car that moves by itself (automotive) to the car that drives by itself (autonomous).