MIT has developed an artificial intelligence camera to see people through the wall 【Full Text】

Technological development For a long time, it seems that it is an unreachable science fiction fantasy to see people on the opposite side through X-rays through the wall, but in the past ten years, the Massachusetts Institute of Technology (MIT) computer science and artificial intelligence experiment The research team led by Professor Dina KatABI of CSAIL keeps us approaching this kind of "seeing people through the wall" scene.

Their new project-"RF-Pose"-uses artificial intelligence (AI) to teach wireless devices to perceive people's postures and movements, and to see and reproduce people's movements and gestures even across the wall. The researchers used neural networks to analyze the radio signals that bounced off the human body, and then created a dynamic stick figure, which could perform walking, stopping, sitting, and moving limbs in synchronization with human movements.

The team said that the system can be used to monitor diseases such as Parkinson's disease and multiple sclerosis (MS), thereby providing a better understanding of the development of the disease and allowing doctors to adjust the medication according to the situation. It can also help the elderly live more independently, while providing more security for falls, injuries and changes in activity patterns. The key advantage of this method is that patients do not have to wear sensors, nor do they need to always charge their own devices.

In addition to healthcare, the team also said that RF-Pose can also be used for new video games, where players can move around the room. Of course, it can also be used in search and rescue tasks to help search and rescue personnel quickly find survivors. Just as mobile phones and Wi-Fi routers have become an important part of today's homes, such wireless technology can power future homes.

The current challenge facing researchers is that most neural networks are carried out by manually labeling data. For example, training a neural network system that can recognize cats requires viewing a large number of pictures, and labeling each picture with "is a cat" or "not a cat". But radio signals cannot be easily marked by humans.

To solve this problem, the researchers used wireless devices and cameras to collect some examples, including thousands of photos, and the people in the photos are performing different activities, such as walking, talking, sitting, opening doors, and other elevators. They then used these images to extract humanoid images and display them to the neural network and corresponding radio signals. This combination of examples allows the RF-Pose system to better understand the connection between the radio signal and the identified target. After the training, RF-Pose can estimate the posture and movement of the target person without a camera, and only receive radio signals reflected from the human body.

Since the camera cannot penetrate the wall, the neural network will not train the data without the wall. What surprised the research team was that the neural network can learn autonomously and recognize the movements behind the entire wall. In addition to sensing movement, the research team also stated that the system uses wireless signals to identify specific people with an accuracy rate of 83%. This ability is very useful for search and rescue operations, and can help search and rescue personnel to understand the identity of a specific group of people.

At present, the model can only output 2D human figures, and the team is working hard to hope that in the future, it can output 3D human figures that can more accurately reflect small movements. For example, it can be detected whether the arms of the elderly shake frequently, so as to determine whether further examination is required. Through the combination of visual data and artificial intelligence, we can better understand the surrounding environment and make life safer and more efficient.