A group of MIT CSAIL researchers has released a new paper outlining a novel new approach to surveilling the movement of people through objects utilizing radio frequency analysis and Wi-Fi signals. Those objects include walls and other obstructions making a body hidden from the human eye. What’s more, the system can also estimate poses and body position with a high degree of accuracy. In fact, it appears to be far more accurate than machine-vision-based systems since it doesn’t require good lighting, line-of-sight, or predictive algorithms. However, it does currently have a lower general accuracy with regard to positioning some body parts. Unlike those other systems, this project – called RF-Pose – only uses Wi-Fi signals. Those typically penetrate walls but reflect off of the human body. Which means that RF-Pose can effectively see and estimate the position of a person through any material which doesn’t reflect Wi-Fi radio signals.
Of course, RF-Pose has its basis in machine learning and other aspects of A.I. Machine vision also plays a role in teaching the software to understand what a person looks like when scanned for using RF signals. It distinguishes each mapped area into sections representing different body parts and once that process is completed, it relies solely on Wi-Fi radio signals and analysis of the thereof to generate a 2D digital skeletal representation of any person or people it detects. Represented in terms of signal travel time, the team behind RF-Pose says it can accomplish that with a high level of accuracy from up to a second away. The distance that represents is obviously going to vary quite a bit depending on what obstacles the signals encounter along the way. The results, as shown in the video below, are impressive, to say the least.
Brought into real-world scenarios, the researchers believe this could serve as a tool for observing at-risk individuals who might be prone to falling or similar scenarios. That makes sense since, if coupled with a sufficiently programmed A.I., the system could be used to automatically contact help. That would be possible even if the person fell while out of sight of monitoring equipment such as cameras. Beyond that, the technology might be a viable solution for more general property monitoring or security. For example, a burglar caught leaving a more traditional camera’s line-of-sight wouldn’t necessarily prevent whatever crime is committed from being captured via RF-Pose. There are likely a lot of other scenarios which this would be useful in too. With that said, as mentioned above, there are still some research and software tweaks that obviously need to be done. The deep learning also needs to be trained up a bit more to ensure it is more accurate than similar systems which rely on RGB images and predictive algorithms.