Volvo has taken to a cutting-edge technique for reconstructing scenes in 3D virtual worlds to help its cars' safety software expect the unexpected.
It's called Gaussian splatting, and it can "create a vast amount of realistic, high fidelity 3D scenes and subjects from real world visuals." Volvo is tasking its spinoff software firm Zenseact to use AI to generate a range of on-road scenarios with this method.
The company notes that objects in each of these environments can be manipulated and moved around to create rare "edge cases" of traffic situations and train driver assistance systems (ADAS). You can see an example scenario below.
For example, you might have a lone pedestrian crossing the street in front of a car in one instance; in another, you could have that same person move quickly amidst a small group of slow-moving pedestrians. These can be used to test how the ADAS software will react in each situation.
Alwin Bakkenes, who heads global software engineering at Volvo Cars, explained that this could give the automaker a way to prevent a far wider range of untoward incidents on the road. "We already have millions of data points of moments that never happened that we use to develop our software," he said. "Thanks to Gaussian splatting we can select one of the rare corner cases and explode it into thousands of new variations of the scenario to train and validate our models against."
This new tech is powered by a supercomputing platform from Nvidia, which Volvo invested in recently. The Swedish marque has been in business with the chipmaker for several years now, integrating its compute solutions into its cars for advanced safety features.
Current ADAS features typically rely on some combination of cameras, radar, LiDAR sensors, and ultrasonic sensors to detect obstacles and signals to navigate roads. The Gaussian splatting method could serve as another tool for Volvo to further enhance its cars' abilities to take over the wheel safely.
Source: Volvo
