What Is Event Detection in Football? (Goals, Cards, Corners)

Event detection in football is the use of computer vision and machine learning to automatically recognise on-pitch events — goals, shots, saves, corners, fouls, cards and substitutions — and time-stamp them so they can be used for highlights, statistics, tagging and analysis. It is the foundation layer of football AI: almost everything else depends on detecting the right events at the right moment.

How event detection works

Models analyse the video feed and first locate the ball, players, officials and pitch lines. A recognition model then classifies what is happening and attaches a timestamp. The result is a structured timeline of the match — not a highlight, but the raw material every downstream product is built from.

What events can be detected

Detection covers the obvious moments and many subtle ones. RISE recognises 38 event classes across its archive, with a core model at 0.76 macro F1 on five primary classes, trained on 9,065 labelled clips. Common football events include:

• Goals and shots on target
• Saves and goalkeeper actions
• Corners, free-kicks and throw-ins
• Fouls, yellow and red cards
• Penalties and substitutions

Why detection is the foundation

Highlights, automatic tagging and live statistics all sit on top of detection. If the detection layer is weak, everything above it inherits the weakness. That is why we say the moat in football AI is the quality of detection on real footage, not the cleverness of the editor on top. The wider point is in How Is AI Used in Football?.

Detection vs decisions

Detecting that a foul occurred is not the same as judging whether it was a foul — that is officiating. Football event detection describes what happened for production and analysis; it does not referee the match. The distinction matters and we draw it out in VAR vs AI: What's the Difference in Football?.

Where it gets hard

• Rare events — a red card or penalty appears far less often than a pass, so there is less to learn from.
• Broadcast artefacts — replays and overlays can fool a model into double-counting or mislabelling.
• Ambiguity — some moments genuinely are borderline, and a model has to express uncertainty rather than guess.

Live vs post-match detection

Event detection can run live, during the match, or after the fact on recorded footage. Live detection powers instant clips, alerts and in-match statistics, and has to be fast and robust to broadcast quirks in real time. Post-match detection can take a second, more considered pass with full context, which often improves accuracy on borderline events. Many production workflows use both: a fast live pass for immediacy, then a cleaner post-match pass that finalises the archive and corrects anything the live system flagged with low confidence.

Who uses football event detection?

Broadcasters use it for highlights and live graphics, leagues use it to build official archives and data feeds, and clubs use it as the raw input to analysis. The same detection layer serves all three, because they all need the same first question answered reliably before anything else: what happened, and exactly when.

Why timing precision matters

An event is only useful if its timestamp is precise. A goal tagged a few seconds late produces a clip that starts in the wrong place; an offside flagged on the wrong frame is worthless. Good detection does not just identify that an event happened but pins down exactly when, to the frame, because every downstream product — clips, stats, graphics — depends on that timing being right. Precision, not just recognition, separates usable detection from a rough guess.

The building blocks of detection

Football event detection rests on three layers working together: object detection to find the ball, players and pitch; tracking to follow them through time; and recognition to classify the action. A weakness in any layer shows up as a weakness in the events. This is why a system that detects the ball poorly will misjudge offsides and goals no matter how good its event classifier is — the foundation matters as much as the headline feature.

How detection is evaluated

Serious football detection is measured with metrics that weight rare events fairly, such as macro F1, rather than raw accuracy that a model can inflate by nailing common passes. It is also tested on real broadcast footage, not clean tactical cameras, because that is where it has to work. When a vendor quotes a single accuracy figure without naming the events, the footage or the metric, treat it as marketing rather than evidence.