Sports Analytics Computer Vision System
Master's thesis project: real-time player tracking, body orientation estimation, and scanning behavior analysis for football matches using state-of-the-art computer vision.
Overview
My master's thesis at Vrije Universiteit Amsterdam, developed at the Sports Intelligence Lab. A lot of statistical analysis and annotation in football is done manually by humans — this thesis automates that process by proposing a computer vision pipeline for two specific statistics: determining the body orientation of a player receiving the ball at a pass event, and counting the number of scanning checks a player performs before receiving the ball.
Technical Architecture
- Player Detection: YOLOv8 for real-time detection of players, with team assignment via HSV color analysis and K-means clustering
- Ball Detection: Dedicated YOLOv8 model trained on custom Roboflow datasets at 2560px resolution
- Multi-Object Tracking: Norfair tracker with Kalman filters for maintaining player identities across frames
- Pass Event Detection: Automated detection of pass events based on ball possession changes and same-team transfers
- Pose Estimation: GluonCV ResNet-152 for extracting 17 body keypoints per player
- Body Orientation: Shoulder keypoint vector analysis classifying orientation into 3 categories (Open, Closed, Half-Open) with 75% accuracy
- Scan Counting: Angular velocity calculation from yaw angles with a 45-degree/second threshold to count rapid head movements
Body Orientation System
The body orientation system classifies a player's facing direction relative to the opponent's goal at the moment of receiving a pass. Using shoulder keypoint vectors, it determines whether a player has an "open" body (facing the goal), "closed" (facing away), or "half-open" orientation — a key tactical metric for coaching staff.
Pose Estimation and Body Orientation
Using GluonCV's keypoint detection, the system extracts shoulder coordinates to compute orientation vectors. The shoulder vector, its perpendicular (forward-facing direction), and the angular difference to the goal determine the player's body orientation category.
Key Results
- Body orientation: 75% accuracy validated against professional analyst annotations on Women's Super League match footage (Brighton vs. Aston Villa, 2024/25 season)
- Scan counting: Established framework for automated analysis, though video resolution proved to be the primary limitation for facial keypoint accuracy
- Published: Semi-Automatic Estimation of Body Orientation in Football at ACM MMSports '25
Technologies
PyTorch, OpenCV, YOLOv8, GluonCV, Norfair, MXNet, Python