Nvidia Teaches Cars to ‘See’
Nvidia, executives says the self-driving car of tomorrow is probably going to look a lot more like today’s models
“Nobody wants to buy a car with a trunk full of PCs or something spinning on top of it,” said Danny Shapiro, senior director of automotive operations at Nvidia, the Santa Clara, Calif.-based chip maker. “You want it to look like a normal car.”
This month it began shipping a hardware and software development kit to customers like Bentley, Aston Martin, Tesla and Rolls Royce
The kit is a crucial part of carmakers’ training of their onboard computers to recognize objects on real-world roads. When they’re fully trained up, something that should take at least another couple of years, the cars should be able to work with other systems to make decisions in real time, like when to brake and when to swerve
The graphics card takes video input from the cars’ cameras and breaks the images down into pixels. It then looks for edges and groups common edges together, looking for patterns that it can identify as parts of objects, like a pedestrian’s legs or a car parked on the side of the road.
The machine then spurts out a statistical probability of what the object in the image is. The method is called deep learning—a subset of artificial intelligence, where layer upon layer of identifier tags are placed on images, which are then further tagged by subsets or groupings: wheels go on cars, cars can be big or small, cars go on roads, etc
The system can, for example, detect the difference between a taxi cab and a police car and will know that the driver has to pull over for the police car if it’s flashing lights at the driver. It can also learn to recognize images it’s never seen before. This is important because if the car sees a moose on the side of the road, and has never seen a moose before, it might not know to slow down or take evasive action if the moose walks onto the road.
But it’s a race against time—as well as other carmakers—to sufficiently train the cars’ computers to recognize and analyze real-world information. “You have to train the car before you ever send it out,” Shapiro said.
Having better data can be a differentiator for automakers, who are all currently building out their own proprietary deep learning databases.
“Self-driving systems are only going to be as good as their algorithms and their data sets,” he said. “If they have much better safety systems it’s going to be a selling feature for that car.”#driverlesscars