Computer Vision

Football Player Tracking with YOLO and Team Colors

A Python football-video pipeline combines YOLO detection, jersey-based team colors and ByteTrack IDs to produce annotated match footage for visual analysis.

Football pipeline combining YOLO detection, appearance-based team grouping and tracking IDs into annotated match video.

Football match demonstration with team-colored player markers, tracking IDs and a ball marker.

Illustrative football pipeline combining YOLO detection, appearance-based team grouping and tracking IDs into annotated match video.

Thirty-second demonstration from the football repository, re-encoded for the website. Team-colored ellipses and tracking IDs overlay a recorded match. The poster comes from this clip; playback frame rate is not a processing-speed result.Detector validation F1 versus confidence on the 49-image validation split. The all-class curve peaks at 0.90 at confidence 0.359; the ball curve is weaker. This chart measures detection, not tracking accuracy.Raw detector validation counts for ball, goalkeeper, player and referee. The background row represents missed detections; the background column represents unmatched predictions. It is not a fifth trained class.

The problem

Turning recorded football footage into a useful visual-analysis baseline requires more than detecting people. Players need team labels and tracking IDs, while goalkeepers, referees and the small ball need distinct detection classes. This project connects those components in one configurable video-processing pipeline.

What I built

YOLO detector training for four classes: ball, goalkeeper, player and referee, using a public Roboflow football dataset.
Jersey-based team grouping that adapts Roboflow's team-classification approach, using pretrained SigLIP embeddings, UMAP and two-cluster KMeans.
ByteTrack IDs through Ultralytics tracking, with an alternative BoT-SORT configuration.
Team-colored player ellipses, tracking labels and a triangle marker when the ball is detected.
Hydra configuration for model, tracker and annotation settings, with annotated video output.

The diagram shows the implemented path from recorded video to the overlay. The gallery clip demonstrates team colors and IDs, while the two charts describe detector validation.

Published detector validation

The repository reports a YOLOv12L result on 49 validation images containing 1,174 labeled objects: precision 0.948, recall 0.880, mAP50 0.916 and mAP50-95 0.675. Player mAP50 is 0.994; ball recall is 0.618, making the ball the weaker class. These are published box-detection results, not measurements of tracking identity consistency, team assignment or processing speed. The evaluation was not rerun for this case study.

The outcome

The implementation turns a match recording into an annotated video that can be inspected with team colors and tracking IDs. The public clip and validation plots make the result concrete, while the weaker ball detection identifies a useful target for further model work.

View the public football tracking implementation.

Where this fits

Useful for scouting products, clubs and software teams building a recorded-football analysis feature. Share a permitted sample clip, camera setup and required overlays to scope Computer Vision Model Development.

Working on something similar?

Discuss a similar project