Turning ordinary football footage into player-level performance insights, using computer vision.
Video not playing? Watch it on YouTube ↗
Grassroots football players have very little access to objective performance data. Professional footballers have analysts, tracking systems and performance reports. Amateur players generally just have their memories of the match.
Use computer vision to turn ordinary football footage into player-level performance insights.
From a single camera, GoalPro detects and tracks every player and the ball, maps their positions onto a tactical view, and builds live match stats.
Detecting players in a single frame is only half the job. To build stats for each player, GoalPro has to follow every player across the whole match — and keep their identity correct the entire time.
In object tracking, there are three main types of error. Because of them, we always need a post-processing stage to make the final tracking accurate and consistent.
The tracker loses a player's identity.
The identities of two players get swapped.
A player's tracking ID incorrectly changes.
Performs the initial tracking and generates tracklets — short pieces of a player's path.
Detects and corrects tracking errors, rebuilds the tracklets and produces the final tracking results.
We spot errors using jersey number recognition and team classification. If a tracklet suddenly shows a different jersey number or team, something has gone wrong — so we split it there. Then we use the same information to work out which pieces belong to the same player and reconnect them.
Result: each player has one complete, correct trajectory for the whole match.
Having something distinctive for each player — here, jersey numbers and team — makes both stages far easier. Without them, post-processing becomes much more complicated and computationally expensive.
The online tracker also gives us information across many frames. So instead of reading a jersey number from a single frame, we use every observation in the tracklet. One blurry or misread frame can't cause a wrong assignment.
Our final tracking accuracy doesn't depend only on the online tracker. A significant part of it comes from post-processing: detecting errors, splitting incorrect tracklets and reconnecting them using reliable player identity information.
From pen and paper to a trained model, one stage at a time.
Before building anything, I needed to know: do people even want this? So I tracked player stats manually with a pen and paper — goals, assists, shots, dribbles, interceptions, tackles, saves and clutch moments.
Doing it by hand took far too long, so I needed to automate it with computer vision. That meant weeks of tagging players and the ball in match footage, often for 3+ hours a day, to create the training data.
I fine-tuned an RF-DETR object detection model on the annotated data, so it can find every player and the ball in each frame of a match.
Next, the model needs to recognise what happens in a match, not just who is where. Here are shots I've tagged — with timestamps and the player who took them — which I'll use to train the model to detect shots automatically.
The next steps to make GoalPro fully automatic: from detecting more of what happens in a match to tracking players without special equipment.
Automatically detect shots, trained on the shots I've already tagged.
A dashboard where players can explore their stats.
Detect other events, such as tackles.
Track players accurately without the custom bibs used for the MVP.
Every match a player plays builds their profile: performance stats, highlights and match history — the kind of data that used to be reserved for professionals.
Whether you run a venue, a league or a team, or you're just curious about the artificial intelligence, I'd love to hear from you.