An institution-grade computer-vision pipeline that reads a tennis player’s biomechanical state from one broadcast camera—and tests whether its decline predicts outcomes

1The thesis and the pipeline

The complement to ball tracking is a body-state instrument. Panopticon follows the biological and psychological engine of the athlete: how the skeleton organizes itself, how the feet load and recover, and how movement changes as a match develops.

The architecture is deliberately linear. Historical broadcast video enters an offline, Zero-Disk pipeline; permissively licensed perception models recover the player; biomechanical features and their uncertainty are extracted; and a separately labeled multi-match exam tests whether those signals predict real events. The production input stays ordinary—one fixed broadcast view. Multi-view hardware appears only in the certification lab.

2Moat A: depth-resolved posture—the signal a single 2D view is blind to

A single camera records rays, not points. Motion along the camera’s line of sight—a trunk leaning forward into depth—can foreshorten to less than a pixel while the true lean grows by tens of degrees. A standard 2D pose estimator is not merely noisy here; it is structurally missing the relevant axis.

Panopticon adds a learned 3D anatomical prior to resolve the hidden component. On clean frames from the current match, median trunk depth-lean is 16.6°, with a 29.6° p95. That establishes the existence of a large posture component discarded by the image plane. It does not certify every reported degree.

Where reality is checkable, the 3D layer’s knee flexion tracks the observed 2D knee angle at Spearman ρ=0.665, or partial ρ=0.60 after controlling for the crop shared by both estimates; a null control remains flat. We label this CONDITIONAL, because it comes from one match and checks only the component 2D can already observe.

The leading builders in human-motion capture independently reach the same practical conclusion: accurate 3D requires multi-view geometry. That is both a defensibility argument and an accuracy warning. Panopticon holds both readings at once. We claim the monocular components that survive—the existence and direction of depth posture—and reserve magnitude certification for the rig.

3Moat B: displacement kinematics with confidence

The recent unlock is not another coordinate; it is a measurement that knows its error. A stride-based odometer removes the failure mode that made single-camera distance look impossible, while the confidence engine reports when position and velocity are trustworthy and when they are guesses.

3.1  The distance dead end, diagnosed and cracked—pending ground truth

The same 100-second clip once produced 638 m, 66 m and 92 m under plausible estimators. The common diagnosis would be calibration bias. The actual failure was noise rectification: taking the absolute value of tiny frame-to-frame position jitter and summing it across roughly 3,000 frames turns zero-mean noise into about 490 metres of phantom travel. A naive odometer cannot attach a valid error bar to that sum.

The fix is a visual analog of zero-velocity updating: measure displacement only between roughly 200 footstrikes. This reduces opportunities for rectified noise by about 15×; each interval contains a real stride around 1.5 m at a signal-to-noise ratio near 10; and a footstrike is the cleanest instant for court-plane projection. The estimator also yields a closed-form confidence interval.

3.2  The confidence engine

Every player position and velocity now carries the Kalman posterior’s model-reported uncertainty: a position standard deviation, a velocity standard deviation and the full tilted error ellipse. During occlusion, that ellipse expands; after re-acquisition, it contracts. The feed can therefore distinguish “the player moved” from “the tracker is uncertain where the player is.”

This is a consequential change for a quantitative model. A downstream consumer can weight, defer or reject a feature rather than silently treating every frame as equally reliable. But the uncertainty is still self-reported and calibration-dependent. Its calibration has not yet been checked against independent ground truth.

4The single-camera ball-bounce localizer—a line-call primitive

A bounce is the one ball event a single camera can place metrically without guessing depth. At contact, the ball lies on the court plane (depth = 0), so its image coordinate can be projected to court metres. That makes bounce landing the primitive beneath line calls and serve placement; it does not make airborne ball trajectories metrically observable.

Detection presence near a bounce is already high—roughly 79–98%, depending on the local window—yet only about one-third of those neighborhoods converts into a precise landing event. The bottleneck is contact-frame dropout: occlusion and motion blur make the ball disappear exactly when the kinematic intersection needs it.

5How we work: we kill our own ideas with math

The process moat is adversarial by construction. Every important result is bracketed by a design review before code and a steelman after the result looks good but before it is committed. Increasingly, that second review uses a different-vendor model to reduce shared blind spots.

This stage catches defects that passing tests do not: a clustering correction applied to the wrong unit, a confound reported from only one side that once inverted a verdict, a fail-open covariance bug that would have emitted “perfect confidence” from corrupt state, and a calibration constant whose name implied the inverse convention. Findings become executable tests or tombstones so the same attractive mistake cannot quietly return.

6The proving ground and the statistics: why N=40, not “more footage”

The fatigue question is a clustered-panel experiment: does within-match biomechanical decline predict the outcome of a point? Frames and points inside one match share player, camera, surface and context. The match is therefore the unit of statistical independence.

Power grows primarily with the number of independent matches, not hours of video. Below roughly N=40 matches, cluster-robust standard errors are not dependable enough for the inference we need. N=40 is an inference-validity floor, not a marketing target and not a guarantee of a positive result.

The proving ground has two distinct arms. The generalization arm asks whether the depth-posture and kinematic measurements recur across players, matches, cameras and venues. The predictive-validity arm asks whether within-match degradation forecasts performance after appropriate controls. Only the second arm addresses the commercial thesis.

7Commercial, IP and capital posture

Shippability is part of the architecture, not paperwork after the science. Models are audited at two independent layers: the repository’s code license and the weights’ training-data license. Permissive code can still distribute non-commercial weights; a clean first layer does not wash a restricted second.

Zero-Disk is operational policy. Video frames stream through RAM or VRAM and are discarded after numeric feature extraction; no raw frame sequence is written to internal storage. Numeric outputs and logs remain auditable without turning the production pipeline into a video archive.

8Status and the critical path, stated plainly

This is early-signal, feasibility-grade evidence built with institutional rigor. It is not a population claim and, honestly, not yet a proven commercial thesis. The value of the program is that its claims are narrow enough to be killed: the sensor faces multi-view ground truth, and the business thesis faces N=40.