EMOTE4D
On-device computer vision that detects falls in eldercare — without sending video anywhere. Skeleton-only processing on Raspberry Pi hardware, running in a live pilot.
The problem
Falls are the leading cause of fatal injury for adults over 65, and the dominant solutions fail for one human reason: pendants and wearables only work if they're worn. Camera systems remove the compliance problem but create a worse one — nobody wants streaming video of their bedroom leaving the house.
EMOTE4D resolves both. The camera's pixels never leave the device. Video is reduced to a skeleton — pose landmarks, not images — on the device itself, and every downstream decision runs on that skeleton. What leaves the home is an alert, not a recording.
Architecture
A pipeline that doesn't trust itself
-
Pose, on the device
A commodity camera feeds an on-device pose estimator. From that point forward the system reasons about biomechanics — joint positions, velocities, torso angle, ground proximity — never raw pixels.
-
Sanity gates between every stage
Compositional vision systems fail when one stage blindly trusts another. EMOTE4D puts explicit gates at each boundary: is this frame usable, is this skeleton real or hallucinated, are these landmarks observed or extrapolated? Phantom poses are killed upstream, before they can look like falls downstream.
-
Two-stage fall confirmation
A fast first stage flags candidate falls from biomechanical features grounded in the physics of falling. A stricter second stage verifies geometry and stillness before anything is allowed to alert — recall first, then precision.
-
Honest failure modes
When the system cannot see — darkness, occlusion, a covered sensor — it says so rather than guessing. A monitoring system that fails silently is worse than none; families are told when coverage pauses.
-
Hardened on production hardware
Every operating number is measured on the deployed device — Raspberry Pi 5 with an infrared-capable camera — in real rooms with real lighting, through multi-day continuous runs. Lab results are treated as hypotheses, not claims.
Where it stands. EMOTE4D is in a live overnight pilot deployment, iterating on the gate architecture through continuous multi-day runs and preparing for in-home family pilots in Manitoba. The engineering record — forty-plus architecture decision records — documents every failure mode found and fix shipped.
Design commitments
What the system promises
- Privacy
- Skeleton-only processing on the device. No video storage, no video transmission, ever.
- Physics first
- Detection grounded in the biomechanics of falling — features that transfer across cameras, rooms, and skeleton formats.
- Affordable
- Commodity edge hardware, not proprietary sensors — built to be deployable in ordinary homes.
Research threads in play
- representation learning
- adaptive computation
- edge deployment
- privacy-preserving vision
- safety-critical ML