The camera is based on time-of-flight technology like that used in Microsoft’s recently launched second-generation Kinect device, in which the location of objects is calculated by how long it takes a light signal to reflect off a surface and return to the sensor.
However, unlike existing devices based on this technology, the new camera is not fooled by rain, fog, or even translucent objects, says co-author Achuta Kadambi, a graduate student at MIT.
“Using the current state of the art, such as the new Kinect, you cannot capture translucent objects in 3D," Kadambi says. “That is because the light that bounces off the transparent object and the background smear into one pixel on the camera. Using our technique you can generate 3D models of translucent or near-transparent objects.”
In a conventional time-of-flight camera, a light signal is fired at a scene, where it bounces off an object and returns to strike the pixel. Since the speed of light is known, it is then simple for the camera to calculate the distance the signal has travelled and therefore the depth of the object it has been reflected from.
However, changing environmental conditions, semi-transparent surfaces, edges, or motion all create multiple reflections that mix with the original signal and return to the camera, making it difficult to determine which is the correct measurement.
Instead, the new device uses an encoding technique commonly used in the telecommunications industry to calculate the distance a signal has travelled, says Ramesh Raskar, an associate professor of media arts and sciences and leader of the Media Lab's Camera Culture group.
“We use a new method that allows us to encode information in time,” Raskar says. “So when the data comes back, we can do calculations that are very common in the telecommunications world, to estimate different distances from the single signal.”
The idea is similar to existing techniques that sharpen up blurred photographs, where several shifted versions of the scene smear together. By placing some assumptions on the model — for example, that much of this blurring was caused by a jittery hand — the image can be unsmeared to produce a sharper picture. The new model, which the team has dubbed nanophotography, un-smears the individual optical paths.
In 2011 Raskar’s group unveiled a trillion-frame-per-second camera capable of capturing a single pulse of light as it travelled through a scene. The camera does this by probing the scene with a femtosecond impulse of light, then uses fast but expensive laboratory-grade optical equipment to take an image each time. However, this femto-camera costs around $500,000 to build.
In contrast, the new nano-camera probes the scene with a continuous-wave signal that oscillates at nanosecond periods. This allows the team to use inexpensive hardware — off-the-shelf light-emitting diodes (LEDs) can strobe at nanosecond periods, for example — meaning the camera can reach a time resolution within one order of magnitude of femtophotography while costing just $500.
“By solving the multi-path problem, essentially just by changing the code, we are able to unmix the light paths and therefore visualise light moving across the scene,” Kadambi says. “So we are able to get similar results to the $500,000 camera, albeit of slightly lower quality, for just $500.”