Tracking players from a broadcast feed

Two matches, one pass: Germany vs Denmark U21 (30 s), then the Spain vs France Olympic final (24 s)

Broadcast left, the same moment in metres right. Track IDs match across both panels. Solid dots are tracked players; hollow rings are officials or people the colour step could not resolve, shown rather than dropped. A dashed box with no number is a player the detector found and the kit-colour test confirmed, but the tracker would not commit to an identity: seen, not identified. Those were previously left off the video altogether, which read as a miss. Full-resolution broadcast overlay, easier to read on a phone: match 1 · match 2. The second match is a different stadium, broadcaster, kits, daylight and frame rate, and nothing was tuned on it. Its kits are unnamed because the kit registry has no entry for that fixture, so the honest labels are Team A and Team B.

Frame inspector — step a frame at a time, see which frames have no calibration, and copy an exact frame reference.

Measured on this window

Where it stands

• Works
  Detection and positions. Players found, teams assigned, and positions in metres on a 102 m pitch, fitted rather than assumed.
• Works
  Accuracy is measured, not asserted. A human judged 317 detections on this exact build. Anything carrying a track ID is a real player 99.7% of the time. Every figure above traces to that review, not to a benchmark.
• Not yet
  It only holds on a wide pitch shot. Across a camera cut identity resets, and a lost track can reattach to a different person. On a zoom, IDs roughly double. On a second match the team labels came out wrong with no warning.
• Not yet
  One window. All of it rests on 27.6 seconds, about 0.5% of a match, reviewed by one person.

Tools