HD Maps for Self-Driving Cars
I used to think that the maps are solved. We have great maps, primarily Google Maps, where roads, buildings, bridges and all other interesting physical objects are mapped. I thought about the future of maps is a series of incremental improvements and ongoing maintenance. Until I learned about HD maps.
Regular maps are not sufficient for autonomous vehicles. They don’t have information about traffic lanes, traffic signs, and lights, position, and height of the curbs. Also, they are useless for localization purposes. They are designed for humans, not for machines. These maps can have location errors up to a few meters.
HD maps (high definition maps) are essential for self-driving cars. They have a high accuracy of object locations, up to 10 cm.
Self-driving cars use HD maps for multiple purposes. It helps to solve the localization problem: figuring out where exactly the car is in the world. GPS can have errors up to a few meters, which is an unacceptable accuracy for driving. With an HD map, the localization software can compare features of surroundings with the map and figure out where exactly it is located.
Then the map can provide valuable information on surrounding objects: where are traffic lanes, which lanes go straight, and which are used for turns only, what traffic signs are nearby, the speed limit on this road, the location of pedestrian crossings and bike lanes, height and location of the curb, etc. All this can be useful for path planning.
Maps are never perfect. Information in the map should contain quality metadata:
- Existence — the probability that the feature exists. For example, when the mapping was done, the mapping software wasn’t sure that a traffic sign was at a certain location due to poor lighting conditions or other faults in sensors of software. The map can reflect that uncertainty by assigning low existence probability to that object.
- Location accuracy in centimeters. It reflects the potential magnitude of error in object location.
- The accuracy of object classification. E.g., the mapping software recognized a traffic sign but is unsure whether it’s “30mph speed limit” or “80mph speed limit”. The map could have metadata for the probability of correct classification, or even a probability distribution across all classes.
The map can also contain information related to the source of the data: which sensors were used to get the information when the map was last updated.
Map errors can be noticed if the vehicles that use the map report any disagreements between their perception and the map back to the mapping service. It is crowdsourcing. This information may be not precise and trustworthy enough to update the map, but at least it can inform quality metadata and guide the mapping company to collect updated information in that location. See HERE self-healing map.
Amount of data to store a map can be enormous. A static base map of San Francisco can take up to 4 terabytes, according to Civil Maps.
Maps can be dynamic. In addition to static data on object locations and features, they can include dynamic traffic and weather information, updated dynamic speed limits in Europe, locations of road works and more.
HD Mapping Companies
Now many self-driving car companies (e.g., Waymo) create their own maps. Also, some companies specialize in mapping services:
There are probably more. Please send me a message if you know about other companies working in this area.
The market for HD maps can be huge. Baidu believes that in the long-term HD maps in China will be a “much bigger business” than Baidu’s search business is today. HERE, a company that was providing mapping and location services for a long time was bought by a consortium of German carmakers Audi, BMW and Mercedes in 2015 for 2.8 billion euros (2.9 billion US dollars).
Links
Autonocast podcast had a great interview with Dr. Sanjay Sood of HERE Technologies. I recommend listening to this podcast if you are interested in the topic.
“HD Maps: New age maps powering autonomous vehicles” — Geospatial World
“Nobody Wants to Let Google Win the War for Maps All Over Again” — Bloomberg
“Edge Mapping: Our Journey from Trains to Self-Driving Cars” — Civil Maps
“The billion dollar war over maps” — CNN Tech
