In an era where technology is advancing at an unprecedented rate, it’s not uncommon to hear outlandish claims about what our devices can do. Some even jest that their WiFi can “read their minds” or that the signals from a cell phone interfere with them and cause them headaches. While that might sound like a stretch, recent developments from Carnegie Mellon University (CMU) might make you reconsider just how far-fetched that idea is.
The Magic Behind WiFi-Based Human Sensing
Researchers at CMU have unveiled a groundbreaking method that uses only WiFi signals to estimate human poses. Yes, you read that right. No cameras, no LiDAR, no radar – just the WiFi signals we use every day. This innovative approach addresses the growing concerns about the cost, power consumption, and privacy issues associated with traditional pose estimation methods.
Originally introduced in 2018, DensePose aimed to map human pixels in RGB images to the 3D surface of the human body. However, the new research takes this a step further. By leveraging three WiFi transmitters and receivers, the system can detect and recover human poses even in cluttered environments. The process involves converting raw signals into coordinates of the human body surface. These signals are then sanitized and translated into 2D feature maps, which subsequently estimate a map. It can be surprisingly accurate and looks similar to the x-ray style maps you might have seen in spy movies or games.
Bridging the Gap Between Image and WiFi
One of the most intriguing aspects of this research is the use of transfer learning. This technique was employed to minimize the discrepancies between traditional image-based networks and the new WiFi-based approach. The results? The WiFi-based DensePose demonstrated a performance that was on par with its image-based counterparts.
Crime Detection: A New Frontier?
This technology’s potential doesn’t stop at mere pose estimation. It could find application in crime detection, because there are already 2/3 WiFi radios within range of most people. While we hope it would be reserved for serious offences, the possibility of using WiFi signals to monitor suspicious behaviours or activities is a possible future method of safely catching crime.
Of course, the hope is that it wouldn’t be used to automatically catch all minor petty infractions and fine everyone. Things like going 2mph over the speed limit, cycling through a red light, vacume cleaning after 1pm on a saturday or singing happy birthday in a restaurant are all technically still illegal – offences many of us might be guilty of probably shouldn’t always be automatically enforced by algorithmic overlords.
Final Thoughts
So, the next time someone claims their WiFi is doing something extraordinary, maybe give it a second thought. With the rapid advancements in technology, the line between reality and fiction is becoming increasingly blurred.
For those keen on diving deep into the technicalities, the full paper is available on arXiv. For a more digestible read, read about it here on Medium. Here is an open source bare bones program that enables you to see how it might work in code https://github.com/Wi-Pose/Wi-Pose. Here is information about how to do it in 3D https://dl.acm.org/doi/pdf/10.1145/3534605 and a another paper about how to do it in 3D https://arxiv.org/pdf/2210.10957.pdf.
Some of the research in this post has come from chatting in the ##machinelearning-general channel on Libera.Chat, you can log in here: https://web.libera.chat/?channel=##machinelearning-general Libera is the biggest open source project chat network, it is home to projects like the Linux Kernel, Ubuntu Desktop and LibreOffice.