Many fall detectors depend on a person wearing and activating a device. I wanted to explore a passive alternative: could a small, inexpensive unit recognise the sound of a fall and call for help without using a camera?
The prototype ran its model on the device. That reduced the amount of audio leaving the room, avoided dependence on a continuous internet connection and made the response faster. Multiple units could cover separate rooms, while two-way communication provided a route from detection to assistance.
The work sat across model development, embedded hardware and the practicalities of care settings. It was developed through the NHS Clinical Entrepreneur Programme and won best talk at the Royal Society of Medicine Patient Safety Conference in 2018.
This was an early lesson in designing around people who may never think of themselves as users of a technology. Reliability, privacy, cost and the path to help mattered at least as much as the classifier itself.
Design constraints
- Passive detection without a wearable or camera.
- Local inference for speed, privacy and resilience.
- Low-cost hardware that could work across multiple rooms.
- A clear escalation path rather than an isolated alert.