How to train a vehicle9/20/19
Autonomous vehicles will soon be commonplace on our roads. As vehicles progress from having autonomous features, such as lane keeping and emergency brake systems, to actual self-driving, they have to become increasingly ’smart’, so the training has to be more sophisticated. This is where companies like Annotell come into the picture.
“We work with machine learning teams in the automotive industry,” says Oscar Petersson, CEO at Annotell, a high-tech start-up company, and one of Volvo’s partners that can be found at CampX.
Autonomous vehicles need artificial intelligence, AI, or machine learning. This in turn needs two things, a machine-learning algorithm and training data.
When you teach the AI how to analyze the environment and how to act on what it perceives, you give it a lot of examples. Volvo takes a huge number of pictures of the driving environment. Then you have to annotate the pictures, that is, describe what’s in them to the AI.
That’s where we come in. Every pixel represents information and we have web-based tools for annotating or classifying this information. When the material is annotated, it’s called training data and that’s what our clients can use to train their algorithms.
This may sound a lot easier than it is. In reality, it’s very complex. The requirements for training data are constantly increasing and the access to high quality training data is a bottleneck in the automotive industry.
To make it even more complicated, no two people will interpret a road environment in exactly the same way. When is there, for example, too much snow in the road for it to be drivable? What pixel of the image marks the road end and the beginning of the roadside?
“The world is complex, and we humans interpret the world very differently. Our goal is to support our clients in agreeing on how to interpret the world, and then transfer that agreement to their algorithms by producing large volumes of consistent training data,” says Oscar Petersson.