At Google I/O, the search giant outlined two new features for Google Maps that help map roads more than ten times faster than possible five years ago. The new feature aims to bring maps packed with useful information to consumers, and Google is giving details on how AI makes two of the new features possible. One way is using AI to teach maps to identify and forecast when people are hitting their brakes.
The feature is leveraging AI to tell when drivers are hitting brakes as part of its routing update that helps users avoid situations that require them to hit their brakes hard, such as confusing lane changes or freeway exits. Google is using artificial intelligence and navigation information to identify hard-braking events, which are moments that result in drivers decelerating rapidly. Hard braking is a known indicator of the likelihood of a car crash.
Google says it believes this update will help to eliminate more than 100 million hard-braking events in routes driven using Google Maps each year. Google trained its AI using machine learning models on two data sets. The first data set has information from phones using Google Maps because phone sensors can determine deceleration along the route. The data is prone to false alarms because the phone can move inside the vehicle independent of the car itself.
To defeat that, Google uses information from routes driven with Google Maps when projected on a display using Android Auto. The search giant admits that’s a relatively small subset of data, but it’s highly accurate because Maps is operating in a stable spot. Google is also working to identify other causes of hard deceleration, such as construction or visibility issues.
Google also confirmed that detailed street maps would expand to 50 more cities by the end of 2021. Again, AI is being leveraged to make that possible. This requires Google to remake its mapping process. It has trained machine learning models to identify and classify features across an index of millions of Street View, satellite, and aerial images starting with roads and then moving to buildings and other features. Its models have been updated to identify all objects in a scene at once, allowing the detection and classification of broad sets of features simultaneously while maintaining accuracy, leading to faster mapmaking.