TerraSensing · Eldercare safety & ambient security
TerraSensing uses ordinary Wi‑Fi signals to work out each person's 3D body position. From that, it is designed to detect falls, track how someone moves over time, and tell whether a room is occupied. It uses no cameras, and residents don't need to wear anything.
A standard Wi‑Fi transmitter and receivers fill the room with radio signals.
As people move, their bodies change how those signals travel and reflect.
Our model decodes those changes into the position of every body joint, for each person present.
When the reconstructed 3D skeleton shows a fall, caregivers are notified.
01The problem
2nd
Falls cause an estimated 684,000 deaths a year, with the highest rates among adults over 60. Source: WHO.
1 in 4
Up from about one in five in 2025. Source: population.gov.sg.
Most people won't accept a camera in a bathroom or bedroom, and stored video is a data‑protection risk.
Pendants and smartwatches only work when they are worn and charged, and they are often taken off for sleeping or showering.
02The technology
TerraSensing is built on WiSPER, an AI model developed by TensorDelta.
The model is trained on about 90,000 Wi‑Fi recordings from a public research dataset, each paired with people's true body positions. Parts of the signal are hidden during training, so it learns to fill gaps.
For each moment, it estimates a 3D skeleton for every person in the room: head, torso, arms, and legs.
A second, smaller model corrects the remaining error in each joint position. This cuts the average error from 7.4 cm to 6.4 cm.
Application software turns the skeleton into events, such as a fall or someone entering a room.
03Applications
01
Monitoring for older adults living alone, in assisted living, or in nursing homes. Caregivers are alerted when someone falls, and changes in how a person moves can show declining mobility early.
02
For homes, offices, and other buildings: knowing whether a room is occupied, and whether someone has entered an area they shouldn't be in.
04Research results
6.4cm
Across 7,824 test frames with one to three people; 4.9 cm when one person is present.
30%+
Compared with previously published Wi‑Fi results on the same benchmark.
Measured on the public Person‑in‑WiFi‑3D test set (7,824 frames, one to three people) using the benchmark's standard evaluation protocol. Published figures are as reported by their authors. Research results from a single training run; real‑world performance may differ.
05Current stage
The model has been trained and tested on Person‑in‑WiFi‑3D, a public research dataset with one to three people in a room. No images are used at any point.
Testing in real homes and care facilities, and building the fall‑detection and alert software on top of the model. We are looking for pilot partners.
06Get in touch
