Football

Detection and tracking

• Players and ball: YOLO. Stock COCO weights (yolov8n) map person → player and sports ball → ball; a football fine-tune with ball, player, goalkeeper, referee classes is better.
• Teams: two clusters of torso colours, kept stable and voted per track. The mini pitch uses each team’s kit colour; teams are “A”/“B” (home/away is unknown).

Pitch calibration

With the Ball + Players + Pitch model and football-pitch-detection.pt (32 pitch keypoints), the worker detects landmarks every 0.4 s, fits a homography with RANSAC and trusts it once two consecutive fits agree. Between fits, camera pan, tilt and zoom are followed with optical flow, so calibration holds while the camera follows the ball. Frames with too little grass (graphics, close-ups, crowd) drop calibration at once.

Events

Event	When
ball_detected, ball_lost, ball_move	Ball presence and movement
possession, possession_change	Nearest player for several consecutive frames
pass_candidate, pass	Release, travel, a different receiver
shot_candidate, shot	Speed spike towards the goal mouth
ball_out	Outside the lines for several frames
corner	Ball settles in a corner arc with a player next to it
goal_candidate	Ball crosses the goal line between the posts (unconfirmed)
goal	The broadcast score goes up by one (read on screen), timed at the goal moment

Goals

One broadcast camera cannot reliably see the ball cross the line: goals come with close-ups, replays and celebrations. EyesOnPlay therefore reads the on-screen score (OCR, about once a second, wherever the scoreboard sits): when it goes up by one and reads the same way twice, that is a goal for that team. The goal is stamped at the moment it happened when that was seen: the ball crossing the line between the posts, or the broadcaster’s “GOAL” graphic. Without a scoreboard (e.g. a clean camera feed) a ball between the posts is only reported as a goal_candidate.

Fields and thresholds: Event reference.