Artificial intelligence has learned to follow people through walls

The system for recognizing moving objects behind obstacles using Wi-Fi echolocation did not appear yesterday, but now its maintenance has been entrusted to a neural network. With new algorithms, everything has changed - a technology has emerged, which is called RF-Pose. She does not "see" the person behind the obstacle, but can recognize who it is and what he is doing.

The principle of operation remains the same, the Wi-Fi signal changes its parameters when passing through obstacles, so if we know the throughput of the wall, then from the rest of the data we can calculate the location of objects behind it. RF-Pose works in two-dimensional space, but with exceptional accuracy, as the AI ​​builds a model of an object and analyzes its behavior, rather than simply measuring the oscillation of radio waves. It looks like a child's drawing of a stick figure, and the system is not able to recognize the face, personality of a person. But where his gaze is directed - without difficulty.

Although AI with RF-Pose does not literally see a person behind a wall, algorithms for analyzing and predicting behavior allow it to paint a very reliable picture. Especially if you give the neural network time and material for training - 100 people participated in the experiment, and by the end of the study, the AI ​​confidently identified them “by gait” in 83% of cases. This approach helps to neutralize the disadvantages of echolocation, to minimize the effect of interference on the recognition of hidden people.

The authors of the development say that with the proper level of skill, the AI ​​will be able to recognize what the criminal behind the wall is armed with, and whether he is ready to attack or is simply hiding. And also - to understand by gestures that a lonely patient in the ward urgently needs help, or to collect dirt on negligent employees who sabotage the work process. Since the technical basis for the RF-Pose application is quite simple, the technology has already been included in the "not for free use" category.