Anonymity DensePose from Wi-Fi.

[Fixxx]

New Tech Can See People Through Walls Using Wi-Fi...​

In order to find a person Diogenes famously relied on a lantern trusting solely in optical recognition methods. However, modern scientists propose using Wi-Fi signals for these purposes. Specifically a methodology developed by three researchers from Carnegie Mellon University utilizes the signal from a regular home Wi-Fi router to accurately recognize not only the location but also the poses of individuals indoors. Why Wi-Fi? There are several reasons for this choice. Firstly, unlike optical recognition radio signals work effectively in darkness and are not obstructed by small obstacles such as furniture. Secondly, it's cost-effective unlike LiDARs and radars which are also capable of performing the task. Thirdly, Wi-Fi is already widely available - just grab it and use it. The key question that remains is how effective this method is and what can be achieved with it's help - let's delve into this.

Methodology For Recognizing Human Poses In Images.​

Let's take a step back and first understand how precise recognition of the human body and it's poses works in general. In 2018 another group of scientists introduced a methodology called DensePose. With it's help they successfully recognized human poses in photographs - solely based on 2D images, without using data on the third coordinate - depth. Initially the DensePose model searches for objects in images recognized as human bodies. These objects are then divided into individual segments which are matched with different parts of the body - each segment is processed separately. This approach is used because different parts of the body move very differently: for example, the head and torso behave quite differently from the arms and legs.

​

Using DensePose it's possible to accurately recognize the poses of human bodies in photos and even construct UV unwraps of their surfaces. ​

As a result, the model has learned to correlate a 2D image with a 3D surface of the human body and obtain not only image annotation according to the recognized pose but also create a UV unwrap of the body depicted in the photo (this allows, for instance, to overlay some texture on it). It's particularly impressive that this methodology confidently recognizes the poses of multiple individuals in group photos including in the genre of "graduation photos" where people stand very close together and overlap each other.

​

DensePose confidently identifies the positions of individual figures in group photos.​

Moreover, according to the images presented in the research and the videos published by the researchers the system adeptly handles unconventional body positions in space. For instance, the neural network correctly identifies people on bicycles, motorcycles and horseback as well as accurately determines the poses of baseball players, football players and even break dancers who sometimes move in very unpredictable ways.

​

The DensePose model excels even with very non-standard poses. ​

An additional advantage of DensePose is that the model doesn't require special computational power to function. When using a GeForce GTX 1080 (not even the most powerful graphics card at the time of the research publication) DensePose recognizes 20–26 frames per second at a resolution of 240×320 and up to five frames per second at a resolution of 800×1100.

DensePose From Wi-Fi: Radio Signal Instead Of Photo.​

Essentially, the idea of the researchers from Carnegie Mellon University was to leverage an existing and well-functioning AI model for recognizing human poses (namely DensePose). However, instead of using photographs as input data for recognition scientists used Wi-Fi signals. For their experiment they set up the test environment:

• Two stands with TP-Link home routers equipped with three antennas: one used as a transmitter and the other as a receiver.
• The scene to be recognized positioned between these stands.
• Camera mounted on the stand next to the receiver router, capturing the same scene that scientists are trying to recognize using Wi-Fi signals.

General scheme of the test setup for recognizing human poses via Wi-Fi.​

Next they launched DensePose which recognized body positions using a camera installed next to the receiver router and tasked it with training another neural network that worked with the Wi-Fi signal received by the router. This signal was pre-processed and modified for more confident recognition - but these are just details. The main point is that the researchers were indeed able to create a new model which confidently establishes the spatial positions of human bodies based on the Wi-Fi signal.

Limitations Of The Method.​

However, it's not advisable to rush with headlines like "Scientists have learned to see through walls using Wi-Fi". To start with this "vision" is quite abstract - the model doesn't so much "see" the human body as it's capable of predicting it's position in space and pose with a certain probability based on indirect data. The actual complexity of achieving any detailed visualization using Wi-Fi signals is demonstrated in another work on a similar topic where researchers experimented with objects much simpler than human bodies - and the results were far from ideal.

​

Visualization of objects using Wi-Fi signal: the less pronounced the edges, the worse the outcome.​

The model built by the researchers from Carnegie Mellon University significantly lags in accuracy compared to the original method of pose recognition in photographs and tends to "hallucinate" quite significantly. The model faces particular challenges when dealing with unusual poses or scenes involving more than two people.

​

The Wi-Fi-DensePose model struggles with non-standard poses and a large number of human bodies in one scene.​

Furthermore it should be emphasized that the test setup configuration in the research was highly favorable: well-known and simple geometry, direct visibility between the source and the receiver, no significant interferences in the path of the radio signal - the scientists created ideal conditions for "penetrating" the scene with radio waves. In real life recreating such a fortunate configuration is likely never going to happen. So, if you've already started worrying that someone will hack into your Wi-Fi router and start monitoring your activities at home it might be a bit premature. If there's anything to fear in your home it's household appliances: for example, smart pet feeders or even children's toys: they have cameras, microphones, cloud connectivity and for robot vacuums - LiDARs that work in darkness and even the ability to move in space.