Artificial intelligence system created to enable self-driving cars to ‘see’ corners
An artificial intelligence system that allows self-driving cars to ‘look around’ the corners in real time can, according to the developers, help prevent accidents.
Researchers at Stanford University in the US have developed a system that bounces a laser beam off a wall to create an ‘image’ of objects that are hidden from view.
The recorded ‘image’ makes no sense to a human being, but with the help of artificial intelligence technology the system can make a visual reconstruction of the hidden representation.
The research was funded by the US government agency DARPA (Defense Advanced Research Projects Agency) and is one of a number of similar technology programs being developed.
It can also be used by soldiers to look around walls, rescue workers looking for people and even in space travel to investigate the inside caves of an asteroid.
Scroll down for video
The systems can one day ‘look around’ self-driving cars in parked cars or busy intersections to not only see cars, but also to read license plates
In addition to the Stanford researchers, the team included experts from Princeton University, Southern Methodist University and Rice University.
The researchers used a commercially available camera sensor and a powerful but standard laser in the new system – similar to those found in a laser pointer.
The laser beam bounces from a visible wall to the hidden object and then back to the wall, whereby an interference pattern is referred to as a speck.
“Reconstructing the hidden object from the speckled pattern requires solving a challenging math problem,” said Metzler.
He said that short exposure times are needed for real-time imaging, but produce too much noise for existing algorithms to work.
A camera uses light scattered from a rough wall, known as a virtual detector, to reconstruct an image of the hidden object. When using a continuous wave laser, the camera records speckles
To solve this problem, the researchers focused on deep learning, a form of machine learning that is better for interpreting large and varied data.
“Compared to other approaches to non-line-of-line imaging, our deep learning algorithm is much more robust to noise and can therefore work with much shorter exposure times,” said co-author Prasanna Rangarajan.
“By accurately characterizing the noise, we were able to synthesize data to train the algorithm to solve the reconstruction problem.”
The artificial intelligence system effectively filters out the noise to create an ‘image’ of what is hiding behind the wall or object.
He said they could do this using deep learning without having to record expensive training data, as would be necessary with traditional machine learning.
“Our imaging system offers uniquely high resolutions and image speeds,” said Stanford University research team leader Christopher A. Metzler.
“These features enable applications that would otherwise not be possible, such as reading the license plate of a hidden car while it is driving.”
It is designed to display small, high-resolution objects, but can be combined with other systems to produce low-resolution images of larger items.
“Non-line-of-sight imaging has important applications in medical imaging, navigation, robotics and defense,” said co-author Felix Heide.
“Our work takes a step towards its use in a variety of such applications.”
They tested their new technology by copying images of 0.4 inch high letters and numbers that were hidden behind a corner.
The research was funded by DARPA, the Defense Advanced Research Projects Agency and is one of a number of similar technology programs being developed
An imaging system was set up about a meter from the wall and hid the letters and they used an exposure length of a quarter of a second.
This one approach produced reconstructions of the real letters hidden behind the wall with a resolution of a quarter of the original image height.
The study is part of DARPA’s Revolutionary Improvement of Visibility by using Active Light-Fields (REVEAL) program, which develops a variety of techniques to display hidden objects around corners.
DARPA says on its website: ‘The REVEAL program aims to develop an extensive theoretical framework to enable maximum information extraction.
“Extracting from complex scenes by using all photon paths and utilizing the different degrees of freedom of light.”
The researchers are now working to make the system practical for more applications by expanding the field of view so that it can reconstruct larger objects.
The research is published in the journal Optics.
WHAT IS DEEP LEARNING?
Deep learning is a form of machine learning that relates to algorithms that have a wide range of applications.
It is a field that is inspired by the human brain and focuses on building artificial neural networks.
It was originally formed on the basis of brain simulations and to make learning algorithms better and easier to use.
Processing large amounts of complex data then becomes much easier and allows researchers to trust algorithms to draw accurate conclusions based on the parameters that the researchers have set.
Existing task-specific algorithms are better for specific tasks and goals, but in-depth learning allows a broader scope of data collection.