Detecting a heartbeat through a wall with IR-UWB radar

Thermal imaging is the standard tool for search and rescue, and it has a failure mode that matters enormously in a burning building: a hot object looks like a person.

A smouldering sofa reads as a heat signature. So does a radiator. So does a body. In smoke, at low visibility, with a two-minute decision window, a false positive costs a responder a trip into a room that didn't need entering.

For Vision-X — our Qualcomm Snapdragon Multiverse entry, which reached the national Top 10 — we built the pipeline around a different signal entirely.

Breathing is motion, and motion is Doppler

Impulse-radio ultra-wideband radar sends extremely short pulses and listens for reflections. Walls attenuate but don't block them. What comes back is a range profile: energy versus distance.

A living person's chest moves. Roughly 0.2–0.5 Hz for resting respiration, sub-millimetre displacement. That motion phase-modulates the reflected pulse. A sofa does not do this.

So the question stops being "is something warm here" and becomes "is something here moving at respiratory frequency" — which only living things do.

The pipeline

1. Clutter removal. The dominant return is static: walls, furniture, the radar's own coupling. Subtract a running average of the range profile, which cancels anything that isn't changing.

2. Range-FFT. Transform to identify where along the range axis the residual energy sits. This gives distance-to-target.

3. Micro-Doppler analysis. For each candidate range bin, take the phase over a time window and look at its frequency content. Respiration shows up as a narrow peak in the 0.2–0.5 Hz band.

4. CNN classification. The spectrogram from step 3 goes to a small convolutional model that separates respiratory signatures from periodic non-human motion.

Making it survive the building

The interesting engineering problem wasn't the signal chain. It was that disaster sites lose connectivity, and a system that stops working when the network drops is worse than no system.

We designed four tiers with graceful degradation:

• Arduino UNO Q at the sensor edge — clutter removal and range-FFT locally, always works
• Mobile — hosts the CNN, aggregates multiple sensor units
• Snapdragon Copilot+ PC — building-scale sensor fusion, plus on-device LLM voice guidance so responders never touch a screen
• Qualcomm Cloud AI 100 — multi-building coordination

Every tier keeps functioning if the ones above it disappear.

Grounding it

None of this is novel research — it's an integration of work that already exists. MIT CSAIL's RF-Pose and WiTrack established that RF sensing can recover human presence through walls. Our contribution was the feasibility study and architecture: what it takes to put that on hardware a fire crew can carry.