Self-driving cars can now see objects around corners

Deep learning based on laser systems can image corner objects in real time, which can be used in the military and rescue in the future

Researchers from Rice, Stanford, Princeton and Southern Methodist University have developed a new imaging method that exceeds previous technologies in resolution and scanning speed, making it more feasible to use lasers to see objects and information beyond the line of sight.

The US military is interested, and its Defense Advanced Research Projects Agency (DARPA) is funding the work, for obvious reasons: NASA wants to use it to image caves, and it wants to be able to do so from space in orbit.

The findings, described in the journal Optica, could one day allow rescue workers to peer into earthquake-damaged buildings or help self-driving cars navigate tricky intersections.

Of course, as things stand, it is still just a scientific project, and even if it has the potential to be applied to the general consumer goods market, it will take many more years.

The earliest peek scheme appeared in 2012, and the breakthrough was to study the time change of the target echo and the transmitted signal after the laser passed through a reflecting surface. But this form of measurement that relies on the projection time requires hours of scanning time to produce an image of a few centimeters.

Other methods have been developed, such as looking at reflected light in an image to infer missing parts, but the latest research focuses on speckles, which are more obvious due to the high coherence of laser light.

This interference pattern model also has drawbacks. To reveal the image hidden in the speckle requires collecting a large amount of spatial information - a process we call non-line-of-sight recording, but the problem is that it is too computationally intensive.

So the researchers used deep learning methods to speed up the analysis. "Image acquisition takes 1/4 second and the resolution is sub-millimeter," explained Chris Metzler, a postdoctoral fellow in electrical engineering at Stanford University and the project leader. "The problem is that the system can only achieve these results by greatly reducing the field of view."

"The speckle coding interferes with the information, and the resolution gets worse as the area gets bigger," said Ashok Veeraraghavan, an associate professor of electrical engineering and computer science at Rice University. "It can recognize a number plate, but it's not ideal for capturing an entire room."

In fact, the two methods mentioned above can complement each other. The projection time of the laser is useful in a wide space, and the speckle analysis helps to read specific information. But in order to avoid interference, all these works require two independent systems to operate lasers of different wavelengths.

At present, the turning experiments are carried out at close range, and the researchers hope to expand them.

"Any method for peeking around corners is still 'lab-bound,' mainly because of interference from sunlight and other issues in real-world scenarios," Veerabhavan continued. "The results published today were achieved under ideal lighting conditions and at a range of one meter. Perhaps trying to emit infrared lasers would be a feasible approach in the future."

There is another intriguing possibility, but that will have to wait until technology advances to shorter wavelengths, which are small enough that most discernible details, including people and cars, are not a problem.