Fixed-Camera 3D: metric 3D from one camera you already have

No LiDAR, no stereo rig. This product turns an ordinary fixed camera into a metric 3D sensor by self-calibrating height, tilt, and focal length purely from the boxes of people already walking through the scene, then reports metric position, speed, and height for anything it tracks. Pure Python (numpy/scipy) math, no GPU, and no model weights involved. Most feed-forward monocular 3D work assumes a moving camera, where the baseline between frames is what makes triangulation possible. Fixed CCTV inverts that: the camera never moves, so there is no baseline, but the person walking through frame is not noise, it is the thing an operator actually wants measured. Scale instead comes from people of known, tightly distributed height acting as a running calibration target.

One verified number

Measured result across the default camera sweep

Explore the height / tilt grid (42 configurations, measured today)

Pick a metric, then click any cell. Each cell is one full run of the real self-calibrate-then-localize-then-speed pipeline against a synthetic scene at that mounted height and tilt (seed 0, 80 calibration walkers, 60 evaluation points, focal length scaled with height the same way the shipped 3-point sweep does). Darker red is worse (higher error or more sensitivity); green is better.

Why the hard case is in the table, not a footnote

The 3 m / 20 degree configuration is a genuinely ill-conditioned geometry for single-camera self-calibration: a shallow viewing angle makes far-field distances fragile even when the calibration residual itself is low. The calibrator's own diagnostic output flags exactly that, in the field the pipeline writes for every run:

"shallow/ill-conditioned geometry, low residual but far-field distances are fragile; treat positions beyond mid-scene as approximate"

We keep that configuration in the headline sweep, in the same table as the two easy cases, rather than dropping it as an outlier. The calibration's own sensitivity metric (sensitivity_m_per_half_deg in src/fixedcam3d/calibration.py) makes the mechanism visible: it perturbs the recovered camera tilt by half a degree and measures how far the median foot point's ground position moves in response. That is a real conditioning probe, not a confidence score bolted on afterward: the optimizer can converge to a low residual while still recovering a tilt that is off by a few degrees, and the residual alone cannot see that failure mode. In the grid above, watch how sensitivity climbs sharply as tilt drops toward 15 to 20 degrees at low camera heights, exactly where position RMSE also blows up, the same conditioning problem showing up in two different measured quantities.

Limitations, stated plainly

• Every accuracy number on this page, including the extended grid, comes only from the synthetic benchmark. Nothing in this repo has yet been validated against real camera footage or a real annotated dataset such as WILDTRACK.
• The camera model is a pure pinhole with no distortion terms and no yaw or roll. Real footage with lens distortion or camera roll will need pre-undistortion.
• Rolling shutter is not modeled, so fast-moving subjects on a rolling shutter sensor will show geometric skew this pipeline cannot correct.
• Ground-contact detection is naive: the bottom-center of a detection box is treated as the foot point, so occlusion can silently produce a wrong ground point with no built-in rejection beyond rays above the horizon.
• The height prior is a single population-mean scalar per scene, not per-subject, so a systematically biased population would bias the calibration in a way the current soft-L1 loss cannot detect.
• Concrete depth-model adapters and the edge or TensorRT deployment path are explicitly not built yet; the interface exists and is unit-tested with no model attached.
• This is research and product-lab work. There is no live customer deployment of this code to reference or report on.