Ground-plane position, height, and separation from a single fixed camera, self-calibrated from the person boxes already passing through frame, no LiDAR, no stereo rig.
Closed source, inquireFixed-camera 3D self-calibrates from ordinary person-detection boxes, with no calibration target and no camera motion required, and turns any upstream detector's boxes into metric ground position, distance, speed, height, and density.
Depth-model fusion math is built and unit-tested behind a pluggable interface, but no concrete GPU or TensorRT depth-model adapter is wired in yet; the project's own README calls that out as future GPU-phase work.
| Mount geometry | Position RMSE | Position p90 | Speed MAE | Height MAE | Calibrated height |
|---|---|---|---|---|---|
| 5 m height, 30 deg tilt | 0.16 m | 0.23 m | 0.08 m/s | 0.06 m | 5.03 m (true 5.0 m) |
| 8 m height, 45 deg tilt | 0.19 m | 0.22 m | 0.08 m/s | n/a | n/a |
This is synthetic ground-truth data with detector-realistic pixel noise, not real footage. The project also discloses an honest failure regime: at 3 m height and 20 deg tilt, calibration converges with low residual but tilt is off by roughly 3 degrees, producing a 4.3 m position RMSE, a case the project's own conditioning probe flags rather than hides.
A pre-mortem risk audit reasons, before spending GPU time, about how real WILDTRACK-class footage would likely break the current method: foot-occlusion bias in dense crowds could inflate calibrated camera height by an estimated 5 to 15 percent, and WILDTRACK's low, oblique camera mounts sit in the same ill-conditioned regime as the 3 m / 20 deg failure case above.
An outsider can download the published benchmark dataset and try the live demo Space to see the synthetic-evaluation behavior described above. The source code is not public, and no real-camera-footage evaluation has been published yet, so the real-data question the project's own risk audit raises cannot currently be checked from outside.