Football Match Analytics

The brief

The client wanted player-level analytics from match video: who ran where, who had the ball, what the chance was worth, and a score that can sit on a player card after full-time.

Doing that by hand does not scale past a handful of matches. The system had to survive amateur pitches, moving cameras and players who overlap.

The challenge

Sports broadcast models assume a TV angle. Real pitches, lighting and bibs do not.

• Detecting players, officials and a ball that is often a few pixels
• Keeping identities stable when people cross or leave the frame
• Assigning teams from kit colour rather than a manual roster every time
• Converting pixels into metres so speed and distance mean something
• Turning positions into events: possession, shots, expected goals, a rating

How we approached it

The pipeline is staged. Detection and tracking run first. Team assignment uses jersey colour. Camera motion is compensated so a pan does not look like a sprint. A homography maps the frame onto the pitch so distances are in metres.

Expected goals and ratings sit on top of that geometry. If the tracking is wrong, the score is wrong - so each layer is validated before the next one is trusted.

What we built

An analysis service that takes match video and writes an annotated output plus per-player metrics: speed, distance, possession, xG and a rating. Team colours and overlays are generated automatically.

The same family of work extends toward a capture unit on the pitch and a cloud job that publishes cards into an app after the final whistle. That hardware path is scoped separately.

Results

The engagement is under NDA. We do not publish league names, player identities, or claimed accuracy against a broadcast product.

In use

The pipeline is used on the client's own footage. New venues need calibration; the analytics path stays the same.