護耆.ai (Wookay)
▶▶ ×4

RGB camera

Ordinary camera
What it reveals
Face, clothes
and your home
RGB inactive

Thermal Sensor

Body height
System alert
Monitoring
Thermal inactive

Radar

Radar · floor position
Movement
Radar inactive
160x120 thermal frames of a fall with skeleton overlay
TODAY
Fall detection
160×120 thermal · Fall #1 · real recording
Skeleton is an AI estimate (illustrative)
Direction AHigh-res skeleton160×120 thermal
160x120 thermal frames with skeleton overlay
  • Full skeleton — joints, posture, gait
  • Lower training cost — existing pose models already work on standing frames
  • Needs a stronger processor (25× the pixels of 32×24)
  • More expensive sensor
Our study: an off-the-shelf pose model found the skeleton on standing 160×120 frames with no training; lying poses still need our own labelled data
Direction BLow-res state tracking32×24 thermal
32x24 thermal frames classified as upright, seated and on floor
  • Low-cost sensor that runs on a local processor
  • No skeleton needed — classifies body state directly
  • Follows the person's state over hours and days
  • Needs broad training data to work in every new home
Our study: 91% of 2-second windows classified correctly (1 person, 1 room)
Low-res works at scale: TaFall (arXiv 2026), 35 people — 98.26% fall detection, 0.00126% false alarms over 27 days in 4 homes
32×24 · automatic
What happened
annotated
UprightSeatedOn floor▼ FallBlank in the annotated row = transitions, not scored
Upright
Seated
On floor
Longest on floor
Falls seen as "on floor"
Windows correct
From minutes to months: time in bed · night-time activity · changes in daily routine · slowing mobility
One person, one room, 20-minute real recording. The automatic row judges every 0.5 s; totals come from it. Accuracy is scored only where annotated. Each segment is judged by a model that never trained on it (3-fold cross-validation).
Loading…