Researchers from the Human Sensors Lab at Carnegie Mellon University (CMU) have published a paper on DensePose by WiFi, an artificial intelligence model that can detect the pose of multiple people in a room using only the signals from WiFi transmitters.
In experiments with real-world data, the algorithm achieved an average accuracy of 87.2 with a 50% IOU threshold.
Because signals from WiFi networks are one-dimensional, most of the previous methods for detecting people using WiFi can only detect the center of a person’s mass and can usually only detect up to one person, Infoq wrote.
CMU’s technique incorporates amplitude and phase data from three WiFi signals captured by three different receivers. This produces a 3×3 feature map that can be transmitted to a neural network that creates UV maps of human body surfaces, which can locate multiple people as well as determine their pose.
The process begins by collecting five channel state information (CSI) samples, which represent the “ratio between the transmitted and received signal wave.” Each sample contains 30 frequencies and is taken from signals sent from each of the three transmitters to the three receivers; the result is two 150 x 3 x 3 raw data tensors, one for phase and one for amplitude. This is converted from a “modal broadcast network” to a tensor of a 1280 x 720 image. It is then processed as if it were an image captured by a camera using the state-of-the-art DensePose pose detection network.
The model is evaluated on a dataset of WiFi signals paired with video recordings of scenes containing one to five people. The scenes were recorded in offices and classrooms. Although there are no annotations on the videos to give a ground truth for the evaluation, the researchers applied pre-trained DensePose models to the videos to create a pseudo ground truth.
In a Hacker News discussion of the work, one user pointed out that in 2020, the IEEE announced the 802.11bf project for WLAN sensors, which is scheduled for release in 2024.
Although the CMU researchers have not published their code or model, the Papers with Code website links to GitHub repositories for three other similar projects.