Autonomous Vehicles Armed with Machine Learning Algorithms
Tom Lombardo posted on May 14, 2018 | | 2023 views

Less Data, More Cognition

In order to navigate around cities, autonomous vehicles (AVs) need up-to-date maps with much more detail than the basic information a GPS unit can store. We're talking several terabytes for just one large city. Clearly, storing all those maps in the car would be unwieldy. On the other hand, keeping the maps in the cloud presents bandwidth issues. It's one thing if a YouTube video pauses to buffer every now and then, but it wouldn't be cool to have your AV just stop in the middle of a street while it waits for a map to finish loading. 

Civil Maps, creator of mapping software for autonomous vehicles, has figured out a way to get the best of both worlds: compress the mapping data so that a local map fits entirely in the car's computer and a new map (say, if you're traveling from one city to another) can be downloaded on the fly. The company just announced that its algorithms have now been optimized to run on Arm processors, an industry leader in automotive embedded controllers.

Livin' on the Edge

Maps for AVs are typically created by vehicles loaded with LiDAR sensors and cameras, which gather information, store it locally, and then ship it to a data center where it's processed and uploaded to a large database.

Engineers at Civil Maps call this "data rich, but operationally poor," due to the inefficiency of moving all that information from the survey car to the data center and then to the AVs. They developed edge-based algorithms that allow the car's computer to handle much of the processing.

Normally, this would take a lot of CPU time and resources, which explains the need for the processing center. But Civil Maps figured out a way to distill all that data into the bare essentials—the "fingerprints"—and to do it in real-time using the car's computer.  

Fingerprinting (Data Compression)

Taking a page from the audio engineer's notebook, the digital fingerprint includes only the "coarse" characteristics of the map, while leaving out the extraneous details, much like sound files are compressed by ignoring subtleties while maintaining the essence of the song itself. Information is captured in six dimensions—X, Y, Z, pitch, roll, and yaw—allowing the driverless car to know its exact position and orientation with 10 cm (4") precision.

Armed with Cognition

Arm processors are among the most widely used CPUs for vehicle control systems, due to their low power consumption, minimal thermal output, and ability to operate in electrically noisy environments. Civil Maps tailored its software to run efficiently on Arm's Cortex series of 64-bit processors - the same ones that provide intelligence to many smartphones, network devices, and advanced driver assistance systems.

Arm's processors incorporate machine learning capabilities that are currently used for speech recognition and computer vision, making these CPUs a natural choice for autonomous vehicle control.

Images and videos courtesy of Civil Maps


Follow Dr. Tom Lombardo on Twitter,  LinkedInGoogle+, and Facebook.

Recommended For You