Features Are Nothing Without Identity: Why Player Tracking Is the Real Foundation of Sports Tech

By Marcell Balogh, CTO / Lead Data Scientist & ML Engineer, Dinkmate.ai

There's a moment every sports tech company eventually runs into: you've built a beautiful feature — shot detection, rally analysis, ball velocity — and then someone asks the obvious question. Okay, but whose shot was that?

If you can't answer that, the feature is a party trick. It looks good in a demo and falls apart the moment someone tries to actually use it.

A stat without an owner isn't a stat

Every piece of sports data only means something once it's attached to a person. A serve speed, a dink placement, a rally length only matters once you can say this happened to this player, in this game. Tracking isn't a layer you bolt on after the "real" features are built — it's the spine everything else hangs off of, and it's easy to forget precisely because it doesn't demo as well as a highlight reel. Strip it out, and every downstream feature — stats, coaching insights, ratings — collapses, because none of it knows who it's actually describing.

Why we built ours in-house

At Dinkmate, we built our own player tracking system for pickleball rather than stitch together someone else's black-box solution. It's the piece everything else depends on, and pickleball has failure modes — doubles players packed shoulder to shoulder at the kitchen line, fast direction changes, a small fast ball — that off-the-shelf models trained on other sports tend to fumble, losing players or swapping identities right when it matters most.

Under the hood: Kalman filters predict where each player should be frame to frame, and the Hungarian algorithm solves the assignment problem, matching new detections to the right existing track. Motion prediction alone breaks down when two players cross paths or one is briefly blocked from view — constant in doubles. That's where re-identification (ReID) comes in: each player gets an appearance embedding, so the system can recognize who it's looking at after an occlusion, not just guess from where they were last seen. Motion tells you where someone probably is; ReID tells you who they actually are.

The takeaway

It's tempting to judge a sports tech product by its feature list. But the real question worth asking about any platform is much quieter: does it actually know who's playing, and can it prove it? Everything interesting sits downstream of that answer.

"When you can measure what you are speaking about, you know something about it." — Lord Kelvin

That's as true of a pickleball rally as it is of physics. Tracking is what turns "something happened on this court" into knowledge you can actually build on.