The system works in real-time, detecting and tracking hidden objects or people using only a smartphone camera.
MIT's Computer Science and Artificial Intelligence Laboratory has developed an algorithm that could help autonomous cars react to hidden obstacles around corners.
The highest-resolution cameras, lidar sensors and radar systems will all fail to identify potential obstacles that could be lurking outside their respective lines of sight, such as a pedestrian that is about to run into the road from behind a parked delivery truck.
Humans have already developed an ability to interpret certain subtle shadows and reflections as evidence of potential danger hidden from direct view. The CSAIL 'CornerCamera' artificial intelligence algorithm aims to achieve an even higher level of situational awareness from the fuzzy shadow, called the 'penumbra,' of objects hidden behind an obstruction.
"Even though those objects aren't actually visible to the camera, we can look at how their movements affect the penumbra to determine where they are and where they're going," says Katherine Bouman, lead author of a paper about the system. "In this way, we show that walls and other obstructions with edges can be exploited as naturally-occurring 'cameras' that reveal the hidden scenes beyond them."
By stacking measurements over time, the system can distinguish a hidden object's speed and trajectory. Such information could be used to distinguish between a hidden telephone pole and a jogger on a collision course with an autonomous vehicle.
The team suggests the algorithm works with a simple cellphone camera, suggesting it could work with integrated cameras on self-driving cars.
Like other forms of machine 'vision,' CornerCameras does have certain limitations. The most obvious drawback is the need for specific lighting conditions, making it less useful when hidden obstacles are shrouded in darkness or if clouds are constantly shifting the ambient light levels. Developers suggest it could nonetheless serve as an important supplemental tool for warning drivers of hidden danger or improving reaction time of autonomous vehicles.