Can AI analyse an amateur match?

Automatic analysis exists and works, on television pictures: high shots, several cameras, professional players in contrasting shirts. The models were trained on that. A club video, taken from four metres up on the touchline, with training bibs and a ball that disappears behind a player twice a minute, is a different problem. Here is what each promise gives today on that kind of footage.

Detecting players: yes

Spotting that there is a person at a given place on the pitch works well, even on amateur video, even at sixty metres. It is the basis for everything else, and it is solid. Errors come from spectators standing behind the line, substitutes warming up, and the referee: the model counts them as players, and you have to tell it where the pitch ends.

Tracking the ball: halfway

The ball is small, often hidden, and looks like a pale boot, a head, or the warm-up ball behind the goal. On an amateur wide shot, tracking loses it regularly and finds it again a few seconds later, sometimes on a different ball. The automatic reframing that depends on it copes by smoothing: when it loses the ball, it follows the group of players, and this is barely noticeable. A possession statistic built on this tracking, on the other hand, is wrong without you seeing it.

Recognising who is who: no

This is the most fragile promise. Reading a shirt number at twenty pixels tall, from behind, in motion, does not work reliably. Systems that “follow your child” generally ask you to point them out at the start and lose them at the first scrum, or rely on face recognition. For young players, face recognition raises a question that goes beyond the technical, covered in filming young players: ask on what legal basis it operates before asking whether it works.

Statistics: the right figure and the plausible figure

Possession, completed passes, distance covered, heat maps: on amateur video, these figures are produced, and they look right. They come from uncertain ball tracking and approximate player identification, and the error accumulates. The danger is how plausible they look. A coach who works on a “38% possession” may be working on a figure that does not exist.

The test that settles it takes one match: count one simple indicator by hand, such as lost restarts or shots, and compare with the automatic figure. If the gap exceeds a quarter, use the system to look at the game, since it cannot measure it. How to count by hand is described in do you need statistics at amateur level.

What AI does well on amateur video

• Finding the phases. Kick-offs, stoppages, half-time, timeouts: cutting an hour-and-a-half file into phases of play is reliable, and saves time.
• Reframing for viewing. An edit that follows the play is pleasant for parents and players, and claims to measure nothing.
• Finding goals and saves. The crowd reacts, play stops, the model sees it.
• Stitching two pictures. Sewing two devices into a panorama and correcting the distortion is a solved problem, which has nothing to do with understanding the game.

What it changes for the coach

Nothing above replaces the gesture that matters: deciding what you are looking for, and finding three situations. AI helps with finding; it does not yet know what deserves to be found in a club match, because that depends on the week's theme and on the team. A tool that lets the coach point to the moment, and takes care of the rest, is more useful today than a tool that guesses.