EAST LANSING, Mich. (WLNS) — Self-driving technology may be getting just a little closer to “seeing” objects in a car’s path the way humans do, thanks to a study led by Michigan State University researchers.
Widespread complaints in recent years about the problem of “phantom braking,” or slamming on the brakes without any visible obstacles present, in self-driving cars have led MSU researchers to dive in and learn and why it’s happening, and how to fix the problem.
Qiben Yan, assistant professor in the MSU College of Engineering, and his research team showed how hackers can deceive the cameras in an automotive vision system, which is composed of multiple cameras and radar using radio waves to gather information.
“We projected lights into the cameras of the vehicle, and the camera recognized this as a false object and hit the brakes; it is surprising how fake things can be created out of nowhere,” said Yan. “We were also able to make an object in front of the car disappear to the camera so that the vehicle couldn’t see an obstacle, and the vehicle hit the object.”
The MSU researchers, including Yan, associate professor Sijia Liu and MSU Research Foundation professor Xiaoming Liu, will use a new National Science Foundation grant, in collaboration with Virginia Tech. They’ll study how the cameras see these phantom attacks and identify ways to map the vision systems more secure and resilient against attacks.
To do so, they’re borrowing from a field relating to human intelligence: neuroscience. They’re looking at studies of how optical illusions can trick human eyes and brains.
“We borrowed some ideas from human perception studies to understand the AI systems behind the vision system,” said Yan. “We know that vision systems can be deceived, and we want to program the AI model to see the environment and interpret the information more accurately.”
Low-level perception indicates that the vision system can see the scene in front of it; mid-level perception recognizes what an object is and can determine how far away it is; and in high-level perception, the system is able to recognize how fast an object is moving and how it interacts with the scene.
“We try to understand the evolution of perception from one level to the next,” said Yan. “Next, we want to develop a defense that ensures inherent security and to enhance the AI model that powers it.”
Yan said the study isn’t just about the vision technologies–“It’s about the overall viability, safety and success of autonomous vehicles in the future,” he said.
