Artificial intelligence (AI) is helping to usher in the age of automated and autonomous driving. Fed with data from a rich sensor suite, AI algorithms can understand and interpret the world around the vehicle and produce the signals needed to safely navigate roads. Part of the challenge with existing approaches involves AI training, specifically the costly and laborious process of data annotation for the machine learning algorithms.brands its This usually involves meticulously labelling elements captured within sensor data images to provide the necessary context for training algorithms. Elements to be labelled include anything from pedestrians and vehicles to lane markings and traffic signs.
There are various ways to tackle this challenge. Some companies are pursuing automated data annotation. Stellantis subsidiary aiMotive estimates that its automated annotation software aiNotate can slash the amount of time it takes to automate an average test-drive sensor feed from a couple weeks to a matter of hours. But this isn’t the only alternative to manual annotation.
Unsupervised learning
A growing number of players are pursuing unsupervised learning. With this approach, an algorithm processes data without predefined labels and derives patterns on its own. Cutting out the element of human annotation could go far in accelerating autonomous driving projects and mastering the long-term challenge around edge cases. That’s the promise from California start-up Helm.ai.
