3D sensor view
Sensor XYZ · meters · orthographicDrag to orbit · scroll or pinch to zoom · arrow keys to rotate · +/− to zoom · Home to reset. Track IDs are reported by firmware; detection and tracking accuracy are unverified. Display range is a viewport setting, not proof of physical coverage or 20 m human-detection range. This is not a map of the whole house.
Vital signs
Firmware estimates · not medical measurementsNo vital-sign samples received in this session.
Motion Signature Lab
Train your own sitting / walking / standing model from observed labels. The browser features and classifier are original; the current TI demo firmware still supplies points and tracks. This is a research prototype, not a safety or medical system.
No posture is treated as truth automatically. No real-world accuracy has been established. Use one consenting adult in an uncluttered test area. Capture and calibration guide
1 · Verify coordinates and repeatable mounting
Measure floor → flat top of the mounting bracket, just before the mounting screws. Add the signed vertical offset from that surface to the AOP antenna center (+ above, − below). These are measured values, not an exact guarantee. Your existing sensorPosition may already transform tracks; verify points and tracks separately. Unknown coordinates block collection.
World mode means the same floor origin and axes as the calibrated sensor transform. Zero roll and upright mounting are required; verify +X sideways, +Y forward, +Z up. Only this lab applies calibration; the main 3D viewer remains in incoming coordinates.
Not calibrated
2 · Select one target and record observed labels
Select the target you are observing. The geometric gate is not a TI point-to-track assignment; nearby tracks are rejected. A session keeps its role. Stop and restart live listening between independent takes. One continuous capture cannot count as two training sessions or mix training and held-out roles.
Optional observed transition annotations are saved in the complete UART capture. Start-up movement is not automatically walking. These markers are human button timestamps, not detected event boundaries; the three-state model does not classify transitions.
No transition event annotation.
No session. Choose a label only after the person is stably in that state; pause before transitions.
No trained model
Waiting for setup
3 · Train and independently evaluate
Collect all three labels in at least two independent training sessions and one separate held-out session. Features are standardized using training data only. Scores are heuristics, not calibrated probabilities. Missing or uncertain observations remain unknown.
No held-out evaluation yet.
4 · Record all available UART output
This independent recorder preserves every received UART byte and every decoded frame while active, including classifier-rejected observations. It is board-only serial output, not raw ADC. Arrival timestamps are the computer’s receipt time, not a synchronized sensor acquisition clock. Recording stops visibly at its memory limit; keep captures short and export each take.
No UART capture. Dataset licensing is pending; exports are private by default.
5 · Save your labeled features
Data stays in memory until you explicitly download it. Reloading loses unsaved work. Export includes derived feature windows, observed labels, calibration and session metadata; treat it as private. No server upload or automatic storage. Import replaces current data after confirmation; only import your own trusted capture files. Export is not a raw UART recording.