Sadji - Real-Time Soccer Player Localization and Tracking

Abstract

Advanced analysis tools leveraging invasive tracking technologies such as gps and manual event tagging has become a global staple in top-tier soccer clubs for enhancing their strategical decision making and insight. These tools rely on precise coordinate data, with their effectiveness significantly enhanced when this data is produced in real-time. With the rapid advancement in computer vision and gpu technology, machine learning models and trackers have become efficient and accurate, enabling real-time production of precise coordinate data using conventional hardware.

In this thesis we will present Sadji, a real-time soccer player localization and tracking system utilizing conventional hardware. We employ state-of-the-art yolov8 deep learning models coupled with multi-object trackers for player detection and tracking. We utilize SuperPoint for keypoint detection and feature matching to produce homography matrices for accurate translation of player coordinates between what the camera observes and a 2D soccer field image.

Through a series of experiments and iterative design choices we gradually improve fps throughput of the system while maintaining a high level of accuracy. Using a combination of interpolation and resolution reduction for input images to SuperPoint, we achieve system throughput speed over 30 fps, while maintaining a comparable position accuracy to that of state-of-the-art gps tracking solutions.