Computer Vision in Soccer

This repository contains the implementation and supporting materials for a project that uses computer vision and machine learning to analyze soccer matches. The primary goal is to extract real-world player positions from video recordings via homographic transformations, and subsequently use these positions and related features in predictive and analytical models.

Overview

• Fine tune YOLO for detecting players in video in real time.
• Estimate field keypoint detections.
• Estimate real-time homographies from detected keypoint.
• Further process group team members via clustering analysis.

Ubuntu-based setup

Create environment and install dependencies.

Place Roboflow training detection data in:

• field : data/00--raw/football-field-detection.v15i.yolov8/
• player : data/00--raw/football-players-detection.v12i.yolov8/

Finetune and Train Models

# train player detection box YOLO
python3 -m scripts.train_detect

# train perspect transformation model
python3 -m scripts.train_keypoints

After train the models

# run real-time inference
python3 -m src.process.real_time

Figure: YOLO-based player detection

Figure: Keypoint CNN pose