Mapping The Maps For Self-Driving Cars (And Drones)

If one doesn’t know where they’re going, sang one Beatle, any road will get them there — but not easily or efficiently. In an age of eCommerce (of two-hour delivery windows, and autonomous vehicles and drones), precision is important when navigating miles, especially the last mile — the last few feet of handing off a package to recipient.

To that end, Sweden-based Mapillary has focused on crowdsourcing imagery for maps. In the last several weeks, it rolled out a platform to crowdsource imagery of various local settings, geographies and street-level data on demand. In terms of the mechanics, the collaborative efforts focus on what is known as street-level imagery, and mapping to satisfy the various needs of users, where getting a literal street-level view of what is where is of utmost importance.

In an interview with Karen Webster, Jan Erik Solem, co-founder and CEO at Mapillary, said that combined efforts on the platform can serve the mapping needs across a range of logistics firms and cities, and can even make autonomous vehicles (and drones) a bit more, well, autonomous.


The Fundamentals

He told Webster that the fundamental problem is keeping maps and geodata databases updated at scale. Over the past decade, he said, there has been an accelerating demand for faster and more accurate updates of maps and street-level views (and no wonder, as eCommerce has made inroads on roads). However, the problem has been that none of the companies that needed the data could get it from their map suppliers.

The platform’s genesis traces back five years, and the newly available marketplace, thus far, covers the U.S. and Europe. There are plans to further expand across a global stage, where, he said, “our contributor network will help get that data in a fast and frictionless way.”

Users download the app and, via crowdsourcing, snap photos with their phones or other image-capturing devices. The images are, in turn, uploaded to Mapillary’s platform, which then segments them according to object types.

“We combine images together, and we machine-generate map data at scale — and we make that data available for all companies,” said Solem.

The companies can then make their own maps so that delivery routes, for example, can be more accurate and customizable, helping them to avoid routing mistakes.

The Pain Points

When it comes to delivery failure (where deliveries are late or nonexistent), said Solem, the most common points of failure can be traced to geocoding or addressing. When someone aims to go to a location on a specific street, the maps must know the latitude and longitude of that location. Delivery companies will typically keep in-house databases of location data, unless they use what is available from Google or other publicly available services.

Even knowing the coordinates is not enough, he said, as delivery companies would ideally like to know where and how the deliveries can be optimized.

Getting Granular On The Road

Picture, then, at the most granular level, the firm that wants to know where parking is easiest so deliveries can be made in the rear of a building, rather than the front. Delivery companies keep in-house databases up to date with fresh inputs so time and money can be saved by finding out where, for instance, parking regulations have changed.

The marketplace — specifically the images as classified by machine learning — exists so firms can, say, put in a request to understand a neighborhood or particular stretch of streets better.

Thus far, according to the company, Mapillary has logged more than 556 million images across a coverage area of 7.5 million kilometers (nearly 4.67 miles). As many as 97 object classes are automatically detected and classified in images (think traffic signs and fire hydrants).

When asked how the company makes money, Solem noted that users can subscribe to imagery on the platform, or map data that the firm generates with the help of that imagery, and pay a monthly or an annual fee. This is dictated by how large of an area to which the users want to subscribe.

On The Road To Scale

The move toward scale is quickened by an inherent advantage in data gathering. As Solem noted, “it costs billions of dollars” for some of the bigger players in the industry to employ fleets of mapping vans to collect data. That gives scale up to a certain point, said the executive, “but we think that our approach gives yet another order of magnitude of scale, … because what we need is just a phone or a vehicle camera on a road somewhere. … If you look around, on most roads, there is someone right there with the tools to do this.”

So far, the densest areas, in terms of coverage, span North America, Europe, Japan and Australia.

The images are contributed, or crowdsourced, globally by individuals, companies, non-governmental organizations and governments. The three biggest user categories are mapping companies, and, according to Solem, the data can be used to fix errors extant in their databases. Cities use the street-level imagery and map features to manage street signs and traffic data.

“When you look around your city, every object that you see, … someone is responsible for maintaining it, for making sure that it is still there and that it is still functional,” he said.

That goes for signage, fire hydrants, utility poles and more. Cities, he noted, take inventory of their signs annually, and look for wear and tear or a need to replace them. The automated detection inherent in the image-capture and classification tied to Mapillary’s platform can save time and money amid those efforts.

The Autonomous Vehicle Opportunity

Autonomous vehicle companies can use the maps to train the perception algorithms used in self-driving cars. That’s an important component of the intelligence that runs within the vehicles, said Solem, and helps it recognize what is in its immediate surroundings.

“One of the key components there is training data. So, if you want to build a solution for accurately detecting pedestrians, for example (and that seems like a reasonable [thing from] a safety perspective), you need to know what pedestrians look like,” said Solem.

In the future, the relationship between mapping and the perception system will evolve.

“So, what the car sees will actually affect the map as well, and keep it updated,” he added.

Travel In The Air

How about drones? A number of parallels can be drawn between autonomous vehicles and delivery by drone. As Solem explained, autonomous vehicles rely on HD maps that can cover everything from bike paths to sidewalks and driveways — “all the areas where there is typically no map today. The recipe there is to collect the data from these robots as they are moving around, and build maps based on that data.”

After all, there are no HD maps of airspaces, but “they will figure out how to move around, based on avoiding things and knowing roughly where they need to go. After a couple of runs, the map will start to form,” as the overlay between perception algorithms and the technology used to generate Mapillary’s mapping data is significant.

“Most of our customers are also contributors to the platform,” he said. Through the marketplace, “our job is to make sure that it’s easy to collect the images, and that the data is generated quickly after the images have been collected.”

He added that, in terms of near-term roadmaps, the focus will be on making sure the marketplace runs smoothly.

“We see a lot of large organizations moving to the platform now, so I expect the volume of data to grow significantly over the next several months,” Solem said.