Testing Radar for Privacy-First Presence Detection

I’m exploring whether a small radar module can answer a useful question — “is someone there, and are they moving normally?” — without ever forming an image or recording audio. This is a note from the early experiments.

Why radar

Cameras and microphones give you a lot of signal, but they also give you a lot of liability. Radar gives you far less, but the “far less” turns out to be enough for some interesting problems.

What radar can do:

• Detect presence
• Estimate range to a target
• Distinguish “someone is moving” from “the room is empty”
• Roughly classify motion (fast vs slow, local vs across the room)

What radar cannot do:

• Recognise faces
• Identify people
• Read text or see objects in any recognisable way
• Hear anything

That second list is the point.

What I learned in the first round

• Range resolution is good enough for room-scale sensing. A small module on a desk can reliably detect motion across a normal-sized room.
• Static objects are not a problem — they fall out of the signal. The radar sees motion, not furniture. An empty room reads as “no motion”, not as “an image of an empty room”.
• Most of the work is in the signal processing, not the hardware. The module gives you raw data. Turning that into “someone is in the room and is moving normally” is a non-trivial pipeline.
• Edge AI matters here. Anything cloud-based would be a privacy regression. The interesting questions — anomaly detection, long-term patterns, calibration — have to run on-device.

Where this goes

The long-term target is a small device a family can put in a relative’s home. The device understands presence and activity. It only surfaces a signal — never a stream of data — and only to the people who need to know.

That’s a long way off. The first round of experiments was about whether the sensing primitive is even viable. So far, it looks like it is.