Today DeepMap is launching a new blog post series called Here Be Dragons. In this series we will explore our love of maps, from the ancient maps on papyrus that guided ocean voyagers, to the digital, highly-detailed maps of today designed to be read by machines.
The phrase Here be Dragons (hic sunt dracones) called out to us as it is associated with unexplored territories on ancient maps. Please join us as we explore mapping topics relevant to the uncharted territory of autonomy. Our first post is guest authored by Brad Templeton and answers a simple question: What are high-definition maps?
What are High-Definition Maps and Why Are They Important?
By Brad Templeton, Guest Author
- A high-definition (HD) map is a highly-accurate 3D map used in autonomous driving.
- HD maps for autonomous vehicles (AVs), such as those created by DeepMap (the host of this series), go far beyond the capabilities of the traditional navigation maps that we use every day, such as Google Maps and Apple Maps.
- “Mapping drives” (by AVs, human driven mapping cars, and ordinary drivers) are able to “see” everything on the road up close and from several angles.
- HD maps are built on many terabytes of data from an AV’s sensors, including cameras and LiDAR.
- The HD map is made from the memory of all prior drives.
- HD maps are important because they provide extra information — and meaning — to AVs
- If a computer is going to take the wheel — and for you to bet your life on that computer’s understanding of the road — you want no less than what HD maps can offer.
* * *
If you drive a vehicle today, you probably make use of “navigation” maps which show the roads and where they go. We used to get these on paper. Now they are often built into our cars, or found on our phones in applications like Google Maps, Apple Maps, and Waze.
These maps have improved a lot by going digital, and now they give us turn-by-turn directions and even figure out the best route based on live traffic. Many people won’t drive without them.
We can drive without them though, but we have a harder time driving without a different type of information about the road, namely all the surface markings and signs which tell us where to drive, what rules to follow, and what it all means.
We don’t call markings and sign information a “map,” but maps and road markings are both extra information about the road that make driving easier and safer, and let us make fewer mistakes. Maps don’t just depict the road, they explain it.
We can drive with just bare pavement or even dirt roads with none of that information, but we don’t do it very well, and we can’t do it very fast.
Now that cars are doing some of the driving tasks — in the hope of soon doing all of it — they need information about the road to do it well and do it fast.
* * *
The very first contest for self-driving cars, known as the DARPA Grand Challenge, kept the route secret until just before the race.
The cars were expected to drive based just on what they could see and a few GPS waypoints.
They did terribly, crawled where they worked at all, and nobody got past seven miles. By the third contest, they were allowed to make a map of the course in advance, and six teams completed the course.
The map they made, though, was not just a diagram of the roads like your old AAA paper map. They realized that any driver — including a computer — has to make sense of the road.
The driver doesn’t just have to figure out where the lane markers are or even decode signs, it has to figure out what all of it means, what every lane marker implies, what each traffic signal means, even when there might be six of them strung over the road. It needs to know where it is, and all the rules it has to follow, and what path to take to get where it’s going.
Those early pioneers, like the winning teams from Stanford and CMU, realized they could put all the things you have to figure out about the road into a better type of map.
The map would show exactly where the lanes were, even where each lane marker was. The map would contain the information on every sign. Rather than expect the car to be able to read every sign, signs and their meaning could be marked on the map..
That way, a car would never fail to stop at a stop sign, and would even know it was coming before the car could see it. The map helped the car understand everything it saw, and know about things even before it saw them well. Stop signs rarely suddenly appear and they almost never disappear. While the car would still need to spot a new stop sign, the map could give it 100% certainly it would never fail to stop at a sign that was already there.
When you’re going to bet your life on some software, 100% certainty is a pretty attractive proposition. As the first self-driving developers filled these maps with more and more details about the road, it was clear they were a new generation of maps. They had a lot more detail, and like the best TV sets of the day, they were called “high definition.”
* * *
All these details let driving systems know where every line and curb and sign is. They also make it possible for a car to reliably figure out its own location on the map, even if GPS and other techniques aren’t working — as they might not in tunnels and urban canyons. GPS is great, but everybody knows it isn’t nearly reliable enough to bet your life on. An HD map can be that reliable.
One big advantage of the HD map is you build it by driving the road in advance in a mapping car, often more than once. That means you see everything up close, and see it all from multiple angles. You can be a lot more sure about what you see than a person or car looking far ahead trying to figure out what the road is like up yonder. You even know what the lane configuration is around the corner to help predict what other vehicles will do.
The mapping car has already been up yonder. More than once. All that data gathered can then be processed with as much computing power as needed. There is no rush, unlike a car speeding down a road at 60 mph that has to be sure it’s figured out what the road is doing before it gets there. Crunch it for an hour on a cloud AI supercomputer if you need to.
After all that processing, humans can look at the map and use our superior intellect to fix any mistakes during a map QA process. Finally, you can send another car, usually with humans on board, to make sure it’s all right. The result is a much better ability to accurately understand the road than is possible in a car trying to see and analyze a different road for the first time.
It’s so useful in fact, that nobody would debate it if it weren’t for two issues. First, roads do change — for example in construction — and you must handle when the road has changed from what the map says it was. Secondly, doing this mapping in advance does take time and money.
If you depend on maps, you can’t drive a road until you have spent that time and money. The first company to try to make a self-driving car, Google, is also the world’s biggest mapping company. By that time, they had already photographed the world and driven every road in many countries for their StreetView project. The leaders of StreetView became the leaders of the car project and applied their expertise.
Since then, like everything done by computers, making the maps has gotten more economical, and you no longer need the resources of a company as huge as Google to do it. As such, an industry has arisen to provide HD maps as a service, or the tools and services with which to build them at scale.
Almost every major self-driving project uses HD maps and views them as essential to the task. (We’ll contrast the views of the one well-known exception in a future article.) Just some of the things you’ll find in these maps include:
- The geometry of each lane, and where the lane markers are, along with what paths you can take between lanes at intersections, merges, and other special locations.
- The exact location of every traffic signal. Ever come to an intersection with dozens of signals hanging in all directions and it wasn’t clear what signal was the one that told you when to stop or go? Computers can get confused too, but a map solves the problem.
- Every traffic rule, like speed limits and all the other rules on traffic signs.
- The locations of obstacles like curbs and barriers to make it 100% sure the car won’t drive into one of these.
- In some cases, the locations of potholes or bad parts of the shoulder, so the car can avoid these things, even if it has to pull off the road.
- All the twists and turns and even the slope of the road, so the car will know in advance if it needs to climb, or take a curve on the inside or outside.
- Clues to help sensors to do better, such as the location of things that reflect radar, or things that might block the view.
- All the places pedestrians might cross the road — or are known to cross even if it’s not officially allowed.
- Special meanings for lanes, such as carpool, bus and bike lanes, or truck or taxi lanes.
- Parking lots and spaces, and where it’s good to pick up and drop off passengers.
- Oh yeah, all those things you need to pick the best path to your destination, just like the maps that people use.
…All this and more — whatever the car wants to be extra sure about. It gets to know all these things before it can even see them. This is what you need if you are going to bet your life, or even just want a driver-assist tool to do a better job with fewer mistakes.
As hinted, everything is not ideal. We’ll discuss the issues of map cost and keeping maps fresh in the future to understand the full picture. One thing though, should be clear. Every driver on the road, human or computer, needs to understand the road and where everything is.
Many hope to make a car that can figure all that out on the fly, letting it drive into fresh territory it has never seen. Every car needs to be able to do that a little, if the map isn’t fresh, but at the same time, any car that can make a map on the fly is a car that can remember what it sees after having driven, and use that memory to do a better job next time.
In essence, a self-driving map, such as those created by DeepMap (the host of this series) is that memory. It would be silly to discard those useful memories, and that’s why pretty much every self-driving team instead organizes them into a map and depends on it to drive.
* * *
About DeepMap: DeepMap is accelerating safe autonomy by providing the world’s best autonomous mapping and localization solutions. DeepMap delivers the technology necessary for self-driving vehicles to navigate in a complex and unpredictable environment. The company addresses three important elements: precise high-definition (HD) mapping, ultra-accurate real-time localization, and the serving infrastructure to support massive global scaling. DeepMap was founded in 2016 and is headquartered in Palo Alto, Calif., with offices in Beijing and Guangzhou, China. Investors include Andreessen Horowitz, Accel, GSR Ventures, Generation, Goldman Sachs, NVIDIA, and Robert Bosch Venture Capital. For more information, see www.deepmap.ai.