In the past we have talked about projects that help us see around corners, but there is another popular goal among researchers: tracking people behind walls. One of the first examples takes us back to the era of the old Microsoft Kinect, but today we discover the RF-Pose system developed by the CSAIL institute at MIT. The system combines wireless signals within WiFi frequencies with a deep learning neural network that processes these signals to generate a 2D model and estimate "pose and location" of a person inside a room.
From rescue operations in fires or collapses to delicate hostage situations, there are plenty of reasons why a security force or an emergency team would love to have a system that allows seeing behind walls. Of course, optics has its limitations, and most of these developments rely on radio signals. One of the main challenges is increasing tracking precision, and a group of experts at CSAIL at MIT decided to apply the firepower of neural networks. The result is RF-Pose:
RF-Pose uses wireless signals within the WiFi frequency band that feed a neural network, which in turn generates a "confidence map" with the received data. That map is sufficient to extract two-dimensional skeletons, regardless of the objects in the room or the lighting conditions. The depth calculation in RF-Pose is somewhat limited (some skeletons disappear when others pass in front), but it can process multiple people simultaneously in a single scene.
Logically, everything comes down to a training issue. The neural network can recognize and isolate a person's walking style, raising the accuracy of pose estimation to 83 percent. In other words, RF-Pose not only manages to determine to some extent what a person is doing, but also who it is. What's next on the list? Moving to a 3D skeleton model that enables the recording of smaller micromovements.
