Demonstrations
—How it works
Both sensors are fixed overhead, so all targets occupy a near-planar floor surface. Rather than estimating a full six-degree-of-freedom transform, a 2D homography maps radar floor-plane coordinates directly to image pixels — sidestepping the lack of vertical variation that made earlier Perspective-n-Point and Direct Linear Transform attempts ill-conditioned.
For detection, RAPiD (rotation-aware, built for overhead and fisheye views) replaced YOLO, which failed on the overhead perspective and on recumbent or stationary subjects. A persistent ghost tracker keeps awareness of a patient after the radar's firmware drops a still target, sustaining the track from low-level point clusters and camera confirmation. A bed-zone anchor holds the track for a sleeping patient who has stopped moving entirely.
The fusion engine runs three independent asynchronous loops — ingestion, detection, broadcast — so a slow inference step never stalls the camera feed or the radar stream.
| Scenario | Frames | Duration | TAR | Mean Dₑ |
|---|---|---|---|---|
| Walking / standing | 1,143 | 143.5s | 99.1% | 103.4px |
| Sitting | 570 | 77.4s | 98.8% | 102.3px |
| Lying down | 247 | 58.7s | 90.3%* | 51.9px |
| Combined | 1,960 | 279.6s | 96.8% | 92.5px |
*Lying-down TAR is conservatively flagged: a bed-adjacent ghost reflection inflated coverage, and the session could not be repeated before this report. Mean Dₑ reflects a systematic radar-vs-camera centroid offset, not spatial error in the fusion zone.
Experiment log
Exp 01–11Seven calibration experiments led to the homography pipeline shown above. Four further experiments document the watchdog and shared-memory work underway this week.