MIT has given a computer x-ray vision, but it didn’t need x-rays to do it. The system, known as RF-Pose, uses a neural network and radio signals to track people through an environment and generate wireframe models in real time. It doesn’t even need to have a direct line of sight to know how someone is walking, sitting, or waving their arms on the other side of a wall.
Neural networks have shown up in a lot of research lately when researchers need to create a better speech synthesis model, smarter computer vision, or an AI psychopath. To train a neural network to do any of these things, you need an extensive data set of pre-labeled items. That usually means using humans to do the labeling, which is simple enough when you’re trying to make an AI that can identify images of cats. RF-Pose is based on radio waves, and those are much harder for humans to label in a way that makes sense to computers.
The MIT researchers decided to collect examples of people walking with both wireless signal pings and cameras. The camera footage was processed to generate stick figures in place of the people, and the team matched that data up with the radio waves. That combined data is what researchers used to train the neural network. With a strong association between the stick figures and RF data, the system is able to create stick figures based on radio wave reflections.
Interestingly, the camera can’t see through walls. So, the system was never explicitly trained in identifying people on the other side of a barrier. It just works because the radio waves bounce off a person on the other side of a wall just like they do in the same room. This even works with multiple people crossing paths.
The team noted that all subjects in the study consented to have their movements tracked by the AI. In the real world, there are clear privacy implications. It’s possible a future version of the technology could be configured only to track someone after they perform a specific movement to activate the system and “opt-in.”
As for applications, it’s not just about spying on you through walls. The MIT team suggests RF-Pose could be of use in the medical field where it could track and analyze the way patients with muscle and nerve disorders get around. It could also enable motion capture in video games — like Kinect but good.