Artificial intelligence for digital maps – International

AI is a hype topic, no company can ignore it anymore. Leen Balcaen, Vice President Product Management SD and HD Maps at Here, is also convinced that Here Technologies has significant advantages through the use of artificial intelligence (AI).

Markt & Technik: Here Technologies also uses AI…

Leen Balcaen: Yes, because our goal is to create the most accurate digital representation of the world around us. That sounds easier than it is – because the challenge in the mapping business is that we are surrounded by a chaotic three-dimensional world created by people from different cultural backgrounds, and we want to translate that into a structured digital map that makes sense and that you can actually feed into software stacks.

We’ve been doing this for years, as you know we use data from many vehicles around the world and collect these data points, which sometimes overlap, sometimes conflict with each other, and from that we generate digital maps. What’s new now, however, is that we increasingly use AI for that.

However, I would like to add that AI is like all technical advances: at first, everyone believes that it will solve all existing problems. And as with all technical advances, this is not the case. AI doesn’t solve all problems either, because it isn’t perfect either. But with AI, we come a little closer to reflecting reality, because we can use it as additional intelligence to improve the information from other sources, achieve results faster and ultimately improve the quality of the information we provide.

Can you give specific examples?

We have been developing our new map creation system, called UniMap, for some time now. To do this, we use our various data sources, combine them with AI and are able to create completely new functions within a few months, whereas previously it would have taken us several years. For example, UniMap uses AI models to automate the processing of 500 million kilometres of GPS data points and sensor data from vehicles per hour in order to extract map features, such as the positioning of road signs in 2D and 3D, validate speed limits and create missing road geometries. This means that it is now possible for changes detected in physical reality to become visible on the map within 24 hours.

Why do we always need new maps, the environment doesn’t change that quickly, not to mention roadworks, and I thought Here passes this information on anyway because it receives this information from the many vehicles travelling around the world?

If you look at normal navigation systems, for example, the driver is told to leave the motorway in 100 metres. That’s no problem for humans, but when it comes to an automated vehicle, this information is not enough. Such a vehicle needs much more detailed information; it must know exactly when to change lanes and when the exit lane begins. Accordingly, the maps for automated driving must be continuously fed with new information, although you are right in that the underlying map is still the same in this case. But for automated driving, we have to change the accuracy and level of detail because the vehicle has to make its own decisions while driving. That’s a key reason why the maps are constantly changing, their state is dynamic, they get more and more detailed and basically more intelligent the further we go. 

Does this mean that these maps contain all the details, at least in the long term?

In principle yes, of course our multi-source approach also helps here, which clearly sets us apart from the competition, who typically only consider one source[SK1] [JS2] , which of course leads to limitations.

The reason why I said ‘yes in principle’ is because I don’t believe that it will ever be possible to create a perfect digital map of the real world, even if you use different data sources like we do and then use AI to extract the best information from the data sources. With this approach, we are leading the way when it comes to the quality of the results, but absolute perfection is not possible.

Does this mean that Here also uses AI to find out which data is correct if it contradicts itself?

Yes, of course, we also use AI to extract contradictory messages. But here, too, I would like to emphasize that AI is not an end in itself. AI is software that is also static in certain respects. AI is not smart per se, it is only as smart as the data you feed it with. In other words, you can also create really stupid AI.

This means that we feed the AI with all kinds of information, e.g. images of a speed limit sign, with signs from Belgium, Germany, Italy, etc. The data comes from all over the world and the AI learns what a speed limit sign looks like based on this data. And yet, in some cases, the AI cannot decide whether it is a speed limit or not.

AI helps us to automate many simple things more, but for certain edge cases, you still need a human brain to look at it and determine what is right and what is wrong. Again, the fact I mentioned earlier plays a role here: humans created the world around us, and a human is by definition illogical, but how can you teach an AI illogical things?

If AI can be used to create maps faster, how often does Here send new maps to its customers?

We already publish map updates on a weekly basis. The update rate depends on the customer’s wishes, some want a weekly update, some a monthly update and others every three months. It depends on the use case.

As you said, not every element in a map changes frequently. For example, a street name, a house number, an administrative area – these things change very rarely. Speed limits, on the other hand, change all the time, for example at roadworks. But we have our streaming services for these changes, which means that a weekly subscription is not necessary for this service, but the data is fed directly into the vehicle.

This means that we are talking about a daily update rate for these streaming updates. This means that, depending on which map element is involved, there are different ways in which we can bring this information to the market and how our customers can use it.

But that was also possible without AI?

Yes and no. In the past, it was definitely very difficult to carry out daily updates. This was due to the processing of the data that we have to carry out. And that’s faster with AI, and with faster processing we can also publish faster, and at a higher quality level.

As you said above, the AI can only be as smart as the data it is trained with, but Here gets data from millions of vehicles, so data shouldn’t be a problem, should it?

Yes, there are many reasons for this. It starts with the question of the extent to which digitalisation is already advanced in the individual regions – there are huge differences here. And even though we’ve had connected vehicles on the road with SIM cards for years, which makes it possible to transmit data much faster, you still have to consider how good the transmitted data is. If you look at the sensors in the vehicle, each sensor has its own advantages and disadvantages and therefore problems. These sensors are not perfect. Then there is the software above the sensors, which is not perfect either. This means that all the sources from which we obtain data have imperfections. We simply receive incorrect information. Another example are road signs. There are cases where not a single car that drives past a specific sign recognises it. This may be because it is very old, it may be because the sun is shining on it, etc. Of course, these sensor technologies continue to develop and improve, but as things stand today, there are still things that they cannot recognise. There is also a second point: a sensor can only detect what is present in the infrastructure. As a human being, you automatically know when you drive into a place that you have to slow down. If there is no speed limit at the start of the town and the town sign is obscured, then no sensor will recognise that you need to slow down.

These are just a few examples, but the combination of the various problems makes it difficult, in my opinion, to create a 100 per cent perfect map in the near future. For some issues, such as speed limits, we have already achieved an accuracy of around 93 per cent, and we are clearly moving towards 95 per cent. But the last five per cent is much more difficult to achieve and requires an enormous amount of effort.

Is there a difference in accuracy when comparing different regions? 

Yes, absolutely. It is particularly good in Western Europe, for example, while Eastern Europe is slightly worse, but here too we can draw on more and more data sources. In the United States, the west and east coasts are very good, while in the central regions, where the population density is also lower, it is slightly worse. India, on the other hand, is now pushing ahead with digitalisation so that more and more data sources are available here too, and the same applies to South East Asia. As I said, digitalisation varies from region to region, and this is also reflected in the accuracy and quality of the maps.

It is clear that an autonomous vehicle needs very detailed information, but there are not yet many vehicles with level 3 and higher in the world. Are there other areas that rely on very detailed maps?

Vehicles with a level 2, level 2+ degree of automation also need information about lanes, for example, as automatic lane changes are a common use case here. Detailed curvature data which is analysed by cruise control, is also critical for ADAS. In many cases, it is also a question of fuel-saving driving, for example for lorries, because if the lorry knows the exact course of the road, it can switch off the engine earlier on downhill gradients, for example, and simply roll. This is important from an environmental point of view, as it reduces CO2 emissions and fuel consumption. We are talking about 3 to 5 per cent of fuel that a truck can save, which is a lot of money for the haulier. Electric vehicles also need a detailed map, because the range changes depending on where the vehicle is travelling, i.e. how many bends, inclines, gradients etc. are on the route.

What are the biggest drivers for creating the most detailed maps possible?

It’s difficult to say because our maps are used in so many applications. But I would say that the biggest driver is driver safety at speed limits, because it’s important to warn drivers in advance. This is part of the ‘European Zero Vision’ goal, which aims to reduce the number of fatalities on Europe’s roads to zero by 2050. The second driver comes from adaptive cruise control, where information such as gradients, bends etc. is also important, and this also applies to lorries.

In addition, detailed information enables better or more attractive visualisations in navigation systems, and we are also seeing increasing demand here.

And in the future, it will be about everything that has to do with autonomous driving. Of course, this is also about visualisation, because this creates trust with the driver if they can follow exactly what the car is doing.

The interview was conducted by Iris Stroh

---

Lesen Sie mehr zum Thema

---