Resources

Highlights

Summary

In this group project, we developed a real-time football analysis system using a fine-tuned YOLOv11 object detection model, capable of identifying players, goalkeepers, referees, and the ball directly from live match footage.

To achieve this, we built two separate components:

  • Object detector: A YOLOv11 model fine-tuned on a dataset of 400 labeled football frames. Using 5-fold cross-validation, we selected the ‘m’ variant for its balanced accuracy and efficiency. Despite great performance on most classes, ball detection remained a challenge due to its small size and visual ambiguity (e.g., penalty spots, players’ heads)

  • Team classificator: K-means clustering applied to the color features (hue and saturation) of the upper half of each bounding box (where the player’s shirt is most likely located)

The system achieves a mean Average Precision (mAP@50) of \(0.827\) across the main object classes and runs at 30 FPS on a consumer-grade RTX 2060 Super GPU.

Pictures

Mean ± standard deviation of AP@50 per class, for each candidate model

Mean ± standard deviation of AP@50 per class, for each candidate model

Normalized confusion matrix for the selected model ’m’, indicating class-wise accuracy and common misclassifications

Normalized confusion matrix for the selected model ’m’, indicating class-wise accuracy and common misclassifications

Precision–Recall curves for the selected model ’m’. A confidence threshold near 0.2 offers a good trade-off between precision and recall

Precision–Recall curves for the selected model ’m’. A confidence threshold near 0.2 offers a good trade-off between precision and recall

Sample prediction made by the selected model ’m'

Sample prediction made by the selected model ’m'

See it in action