I built a prototype that tracks activity without a camera. It uses short-range radar, rough depth sensing and audio level only, with no recording, to spot presence, falls, sudden events and long inactivity. The point was to keep the space private while still feeding useful events into home automation. Here is the hardware, the logic and the bits that matter when you actually try to make it work.
Start with one room and one problem. A bedroom or bathroom is the obvious place. Use an mmWave radar module for presence and motion, a time-of-flight depth sensor for posture and distance, PIR for basic occupancy, and a mic in level-only mode for sound events. Run the fusion on local compute, such as an ESP32 or an nRF54 series chip. Combine radar movement, depth changes and sound spikes to cut false positives. Report labelled events only: presence, fall, loud incident, prolonged immobility. Send those to your home automation hub, not to a cloud recorder. During development, keep diagnostics light and log counts rather than raw data.
The sensor mix matters. Radar modules from Vayyar or other mmWave vendors pick up motion and micro-movement in the dark and through light obstructions. Time-of-flight depth sensors give coarse posture data, which helps separate sitting from lying down. PIR is cheap and low power for simple occupancy. For fall detection, a single sensor is weak. Radar can show rapid displacement, the depth sensor can show body change, and the sound level can jump. Put those together and the false alarm rate drops. Keep the automation simple: turn on lights, call a contact, trigger a siren, or open a voice channel after a confirmed event. Test it in the room where it will live. Sensor position and threshold tuning matter more than swapping modules.
Privacy is the reason to do this in the first place. Raw sensor streams should stay on the local device. Convert audio to a single amplitude value and discard the samples. Use on-device models or rule engines so the hub only sees events and small diagnostic counters. Short retention and log rotation help. Use TLS for notifications and keep private configuration on local storage. Do not use cloud video or audio storage. That keeps the system acceptable in rooms where cameras are a bad fit. It also works in the dark and keeps network traffic small. The trade-off is less detail, so conservative alerts and sensible sensor fusion matter.
I have used this sort of setup in prototypes built around alert-on-unusual rather than constant monitoring. That covers fall alerts, remote check-ins that report presence or long inactivity, and privacy-first room occupancy for heating and lighting control. In practice, the small things matter: angle the radar down a little so it sits on human height, place depth sensors where furniture will not block the view, and set sound thresholds so normal household noise is ignored. Cheaper integrated mmWave modules should keep improving, edge ML on small MCUs should get better, and non-camera device profiles for home automation hubs would save a lot of faff. The simple part is the idea; the useful part is keeping it local and making the alerts behave.
