The problem
Coaches and analysts need usable evidence from a full match, but manually reviewing and tagging 90 minutes of footage slows down feedback and makes consistent analysis difficult.
The approach
A YOLOv8 detection pipeline and OpenCV processing layer follow players and the ball frame by frame. PyTorch models classify match events, while a FastAPI service makes the resulting clips and metadata available to the product experience.
Implementation
The system processes match footage, maintains identities through play, detects important events, and indexes the relevant moments so analysts can move from a match to tactical review without tagging every sequence by hand.
What we learned
- Sports tracking becomes useful when detections are connected to events and coaching questions, not presented as boxes alone.
- Broadcast footage changes constantly, so identity continuity, occlusion handling, and clear confidence signals matter as much as raw detection accuracy.