SPORTS VIDEO · COMPUTER VISION · MEDIA AUTOMATION

Finding active play in fixed-camera tennis recordings.

We built a processing pipeline that tracks both players, maps their movement to a normalised court and identifies serves and rallies so long match recordings can be condensed around meaningful play.

Product: Tennis video intelligence
Sector: Sports media
Engagement: Computer-vision processing system
Status: Delivered prototype

GEOMETRY

Court-aware

Camera positions projected into a common court space

SEGMENTATION

Player-led

No dependency on continuous ball tracking

OUTPUT

Explainable

Event timelines and visual debug evidence

What the prototype established

A court-aware processing pipeline that produces active-play intervals and an explainable timeline from fixed-camera tennis video.

Player behaviour remains the primary signal when the ball is not reliably visible.
The resulting tracks can support later heat maps, coaching statistics and point search without rebuilding the base analysis.

THE PROBLEM AND HYPOTHESIS

Movement is not the same as tennis, and the smallest object is not always the best signal.

A player collecting a ball may move more than a server preparing to begin a point. Generic motion detection therefore retains the wrong material. The tennis ball is a useful signal when visible, but from a single fixed amateur camera it is often only a few pixels, blurred or completely occluded.

The system needed to model the structure of the sport. Player positions become more meaningful when projected onto a consistent court. Serve preparation, court-side identity, speed and temporal transitions can then be combined into a conservative state machine.

Tennis footage comparison showing high motion between points removed and lower-motion serve preparation retained

COURT AND TIMELINE MECHANISM

Stable court geometry made imperfect detections useful.

The combined mechanism projects player locations onto a calibrated court and turns their behaviour into an inspectable active-play timeline.

01 · Test stage
Calibrate fixed camera
02 · Test stage
Register court geometry
03 · Test stage
Project player locations
04 · Test stage
Track player behaviour
05 · Test stage
Segment serve / rally / idle
06 · Test stage
Export explainable timeline

ENGINEERING DECISIONS

Using court geometry when the ball could not be trusted.

Use the court as a coordinate system

Homography makes player position and velocity rules interpretable across perspective. A fixed installation turns camera registration into controlled configuration rather than repeated inference.

Computer vision

Avoid making ball detection a hard dependency

Ball evidence can strengthen a decision when available, but player geometry and temporal behaviour remain the primary signal. The pipeline therefore continues to operate when the ball is blurred or occluded.

Model strategy

Separate analysis from rendering

Detection produces a structured event timeline. FFmpeg later extracts and joins intervals. Thresholds and editing policy can be adjusted without rerunning every expensive model stage.

Media processing

VALIDATION LIMITS

The prototype retained clear operating assumptions

The method assumed a fixed camera and required court calibration. It demonstrated an explainable active-play timeline without claiming reliable continuous ball tracking across arbitrary footage.

RESPONSIBILITY