Researchers propose a system using artificial intelligence to observe hidden objects around corners.

2020-01-17

Researchers from Stanford University, Rice University, Princeton University, and Southern Methodist University have published a study in the journal "Optics" proposing a system that can generate images of turning points with high resolution and speed. According to co-author Felix Heide, the system can distinguish sub-millimeter details of hidden objects from a distance of 1 meter, making it suitable for identifying license plates on hidden moving vehicles and personal badges worn by pedestrians. Heide also stated, "Non-line-of-sight imaging has important applications in medical imaging, navigation, robotics, and defense." Their method is not the first to construct digital images to reveal what is about to happen. Last year, scientists from Intel Labs and Stanford University used multiple sets of speakers and off-the-shelf microphones to capture the time of sound wave echoes, enabling an algorithm inspired by seismic imaging to generate images of hidden objects. Additionally, MIT has detailed a camera that reconstructs scenes outside the field of view by analyzing shadows. What sets this latest imaging system apart is its use of a commercial camera and a laser light source similar to a standard pointer. The beam reflects off visible walls onto hidden objects, then reflects back onto the walls, creating a unique light scattering pattern known as speckle pattern, which encodes the shape of the hidden objects. To reconstruct the objects from the speckle pattern, the researchers used an artificial intelligence algorithm trained on synthetic data with feature noise, eliminating the need to capture experimental training samples. Importantly, it can operate with shorter exposure times required for real-time imaging, as co-author Prasanna Rangarajan explained, "Our deep learning algorithm is much more robust compared to other non-line-of-sight imaging methods." The researchers tested their technology by attempting to reconstruct images of letters and numbers hidden in a corner 1 meter away from the wall, with a height of 1 centimeter. Using a quarter-second exposure time, this method produced reconstructed images with a resolution of 300 micrometers. The development of this system is part of the US Defense Advanced Research Projects Agency's efforts to utilize active light fields to revolutionize visibility. They claim that it can be combined with other imaging systems to produce low-resolution, room-sized reconstructions, which could enable future autonomous vehicles to "see" parking spaces or busy intersections, or assist satellites and spacecraft in capturing images from caves on asteroids.

RECENT AI NEWS

RECENT AI TOOLS