Seeing Through Walls with Wi-Fi: A Developer's Look at WiFi-DensePose
Imagine being able to detect a person's pose—whether they're sitting, standing, or waving—without a camera, without LiDAR, and without any on-body sensors. Now imagine doing it through a wall using nothing but the Wi-Fi signals already floating around your home. That’s the intriguing promise of the WiFi-DensePose project.
It sounds like something from a spy movie, but it’s an open-source research project sitting on GitHub right now. It taps into the subtle ways our bodies disturb Wi-Fi radio waves to infer a human pose. For developers, it opens up a weird and fascinating new sensor modality that’s privacy-conscious, works in the dark, and doesn’t require expensive hardware.
What It Does
In simple terms, WiFi-DensePose is a deep learning model that takes Channel State Information (CSI) from Wi-Fi signals as input and outputs a 3D human pose estimation. CSI data is essentially a fine-grained snapshot of how radio signals are being distorted as they travel from a transmitter (like a router) to a receiver (like a laptop). When a person moves in the environment, they affect these signals in specific, detectable ways.
The model is trained to recognize the patterns in this signal distortion and map them to a standard human skeleton model, creating a real-time, keypoint-based pose—all without optical data.
Why It’s Cool
The obvious "wow" factor is the through-wall capability. Because Wi-Fi radio waves penetrate most interior walls, the system isn't blocked by obstacles that would completely foil a camera. This makes it interesting for applications in privacy-sensitive areas where you might want monitoring but not recording, like in bathrooms or bedrooms.
From a technical standpoint, the cleverness is in the repurposing. It uses the existing Wi-Fi infrastructure (or a simple setup with a few inexpensive routers and receivers) as an ad-hoc radar system. The implementation builds on top of DensePose, a computer vision system for mapping human pixels to a 3D surface, but swaps the visual input for RF data. It’s a brilliant example of applying computer vision concepts to a completely different data domain.
Potential use cases are thought-provoking:
- Elderly fall detection that works in any room of a house.
- Smart home automation that responds to your presence and posture, not just motion.
- Privacy-preserving security for detecting intruders without cameras.
- Interactive gaming or fitness tracking without the need for a console camera.
How to Try It
Ready to experiment? The full research code and pretrained models are available on GitHub.
- Head to the repo: