AI Is Putting Digital Mapping Streets Ahead

AI is Putting Digital Mapping Streets Ahead

By Pierluigi Casale, Chief Data Scientist, TomTom

Pierluigi Casale, Chief Data Scientist, TomTom

Autonomous cars won’t be a reality until we map millions of miles of road in every season, and in every driving and lighting condition. That’s because vehicles cannot rely on their sensors alone: they need accurate 360 images of the road in all conditions to ensure that the car can make perfect sense of its environment and anticipate what’s beyond the next bend or on the other side of that truck. So, what do we do when a job is beyond a human’s capabilities? We turn to artificial intelligence (AI), which will play a fundamental role in creating the driving experiences of tomorrow.

Turning day into night

While it’s impossible to take photographs of every few feet of the road around the world in every possible condition, we can train AI models to recognise rain, fog and snow; darkness and daylight – and then apply these conditions to the images we collect on the road.

We do this with a technique called Generative Adversarial Network where, in essence, AI creates ‘fake’ images and then challenges itself to see if it can distinguish between the fake and real images. By playing and replaying this game, the algorithm rapidly improves until it can automatically create believable, highly accurate images that were never actually photographed in the first place.

At TomTom, we use this technique to create all-weather imagery of 400,000 kilometres of road, so we know how to turn the sun into showers and day into night. But while AI provides an elegant solution to the enormous challenge of mapping the world’s roads in every conceivable condition, we still need to keep the actual images of the road and the physical environment up-to-date.

The wisdom of crowds

One area that shows particular promise in building the maps of the future is through crowdsourcing.

But while AI provides an elegant solution to the enormous challenge of mapping the world’s roads in every conceivable condition, we still need to keep the actual images of the road and the physical environment up-to-date

For example, people can use our TomTom apps to send pictures of where reality doesn’t match what we have recorded in our maps, alerting us to send someone out to check and accurately record it.

It won’t be long before crowdsourcing becomes commonplace in the digital mapping industry, just as it has for online reviews and much else. The challenge that the technology industry faces with crowdsourcing is that people are becoming more wary of sharing their data with third parties. With mapping, it’s not only worries about data security but also that many online maps are ad-funded. How can we trust that these service providers will always send us the quickest route or whether they’ll send us on a detour that takes us past one of their advertiser’s drive-through restaurants?

Free, ad-funded maps certainly have their place, but when it comes to crowdsourcing it’s the paid-for, ad-free mapping providers that have the advantage of trust. People know that if they take images of the road and send them to a company that doesn’t advertise, they will be actively taking part in a community that helps everyone – much like Wikipedia.

Of course, we can’t just rely on the community, because not every road sees a lot of traffic. That’s why the AI team at TomTom also crawl the web daily to find announcements about changes to infrastructure, new building developments, and so forth. This is just one example of how we detect changes and improve maps by using data from multiple trusted sources.

Keeping data secure

But even ad-free mapping providers have to reassure the public about data security, especially given the huge volumes of information they hold. The nature of AI means that all data has to be collected in a single place, and that obviously makes it vulnerable to hacking. Educating data scientists on security and privacy concerns is critical.

It’s vital that mapping technology companies can completely anonymise data and can’t identify anyone based on their location or journeys. With data, scientists naturally hungry for as much information as they can get, privacy and data security must always come first.

One way to solve this challenge is through privacy-aware machine learning, where AI algorithms learn from anonymized raw data. Once trained, the models can be shared, and you can continue to train and enhance the full pool of shared models with new ones. And so the continuous virtuous cycle of learning continues.

Artificial intelligence may be the key technology to delivering the future of driving, but we all have a role to play in responsibly feeding AI’s insatiable appetite for data. The race to autonomous driving and a safer world, free of congestion and emissions, has begun.