MIT trains AI to track people's movements through walls

In 2013, an MIT team found a way to see through walls using Wi-Fi radio signals, and in 2015 the technology was advanced enough to distinguish and track individuals. Now, led by Dina Katabi from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), the team has hooked the latest version, called RF-Pose, to an artificial intelligence (AI) neural network as a way to teach it to deduce a person's postures and movements even when completely hidden.

RF-Pose works by beaming very low-powered signals through walls and other obstacles and then processing the reflected radiation to build up a 3D scan of the area. To create a dynamic stick figure that can move and position its limbs in the same way as the subject, the MIT team used a neural network that's capable of learning by comparing thousands of radio scans and photo images of the same examples of a movement or posture, including walking, talking, sitting, opening doors, and waiting for elevators.
After training, the images were dispensed with and the system relied purely on the radio data. According to the team, RF-Pose was able to generalize its knowledge and make identifications with a surprising degree of accuracy, despite not having cameras for making comparisons. Using wireless signals, it could even identify people out of a line-up of 100 individuals with a success rate of 83 percent.
The researchers see RF-Pose as having a wide range of potential applications. Far from being a spying device, it could be used to monitor patients with Parkinson's disease, multiple sclerosis, and muscular dystrophy to gain a better understanding of the diseases. It could also be used to keep an eye on elderly patients to provide them with a more independent life, and to aid search and rescue teams in not only finding survivors, but identifying individuals. It can even provide early diagnosis for certain neuromuscular conditions.

However, the team emphasizes that the scans would only be conducted with the subject's consent with the data anonymized and encrypted. In addition, scans would only be initiated when the subject makes a preprogrammed gesture.

"We've seen that monitoring patients' walking speed and ability to do basic activities on their own gives health care providers a window into their lives that they didn't have before, which could be meaningful for a whole range of diseases," says Katabi. "A key advantage of our approach is that patients do not have to wear sensors or remember to charge their devices."

The research will be presented at the Conference on Computer Vision and Pattern Recognition (CVPR) in Salt Lake City, Utah.

The video below shows RF-Pose in action.