How we make our HD maps

At TomTom, our goal is to map the entire world for autonomous vehicles to drive safely. For this to happen, we must map tens of millions of kilometers of road every year. Making HD maps is not only a challenge because of the sheer amount of data that needs to be captured and stored, but also because of the intrinsic accuracy requirements.

To tackle this challenge, we rely on automation to scale the production pipeline. Using deep learning and machine learning algorithms, we are training AI models to perform tasks more accurately and at a much higher speed than humans.

Machine learning is all about learning from data and using the right kind of data in the right way in order to solve a certain problem. Its end goal is to gain the ability to make predictions based on previously unseen data.

Let’s take the example of traffic signs. To classify them, you would first need to show a machine learning algorithm many images with variations of traffic signs, from which it could learn a model for how each sign looks like and what makes it distinguishable. Once the algorithm would encounter a never-before seen image of a stop sign, it would be able to recognize it.

A similar approach is required to detect traffic signs. However, instead, you would need several images of road scenes with a rectangular box (or a mask) on top of each sign in order to teach the algorithm how a sign looks and where it usually is located in the environment.

This process extends to many other scenarios that are relevant to autonomous driving. One other example is labeling road geometry, such as signposts and road markings and integrating this information in our HD maps. Making accurate and consistent annotations is key to a map that knows every part of the road environment from every angle and condition.

Using AI is not only helping us achieve what would have otherwise been impossible in mapmaking but also accelerate the progress towards Level 5 autonomy. Let’s look at four important reasons why AI is key to creating HD maps.

Manual annotation simply does not scale. When the task at hand is to detect 30 types of lane dividers, some of them, such as solid and dashed lines, occurring frequently, and others only found in certain countries, it takes even experienced human map editors a considerable amount of time to annotate each divider correctly.

When done by a machine learning algorithm this process is sped up considerably, allowing for ten times more annotations to be made in the same time it would take someone to go through this manually.

For HD maps to be reliable, they must offer an objective and consistent representation of the road at all times. When humans are in charge of interpreting and labeling data, there is always a margin for error due to our inherent bias. This is because instructions can be interpreted in slightly different ways by different annotators when something is not 100% clear.

Let’s take the example of an image taken by a MoMa car, which needs to be processed for recognizing elements that will then be fed into the HD map. When an element in the picture is open for interpretation – be it a road sign or a lane divider – different annotators might label it differently, leading to clusters of different types of annotations due to several ways of interpreting the same element. When a machine learning model takes over interpretation is more consistent and stable, leading to a cleaner set of data.

Although high-quality manual annotations are often crucial for training robust machine learning algorithms, and human operators are an important stronghold for quality checking our HD maps, the key is to find the sweet spot where the minimal amount of human intervention is required and quality is kept within requirements.

For example, at TomTom, we use AI to aid manual annotations. Human operators are presented with annotations generated automatically by the machine learning model and their task is to check them and correct any mistake.

The corrections are then used to train an improved machine learning algorithm that will be less prone to mistakes, thus closing the virtuous circle that will lead to the desired sweet spot. With time, human operators can be reassigned to more meaningful and less tedious work.

For HD maps to play their role successfully, they need to reflect changes that happen on the road as close to real-time as possible. AI helps us reduce processing time as much as possible in order for updates to reach the map faster.

Using AI also leads to reduced costs. Once trained, the machine learning model can perform tasks efficiently. When the choice comes between training an algorithm or hiring several hundred people to process the same amount of data, it becomes clear how significant cost reductions can be if the technology is used instead of or to complement the work of human annotators.

The safest autonomous vehicles are those with HD mapping capabilities. It helps them see further than their sensors and with far greater context. It results in more accurate driving and faster, better decision making when the unexpected occurs.

For TomTom, using AI to create HD maps leads to peak accuracy at the lowest cost possible, enabling us to help the automotive industry reach higher levels of automation, faster.

So let’s take a photo of 5th Avenue in New York City today and compare the same shot in 10 years' time. I have a feeling we’ll be as surprised by the changes as we were almost 100 years ago…