Joavku: Real-Time Team Assignment using Visual Data

Abstract

This thesis presents Joavku, a lightweight and interpretable system for performing automated team assignment of football players using only visual data. Unlike many existing systems that rely on custom-trained machine learning models or external tracking technologies, Joavku utilizes a novel color-based classification method that identifies the dominant team color in a player’s kit. By isolating key visual regions of each player image and analyzing them in the perceptually uniform CIELAB color space, the system achieves high classification accuracy without requiring manual annotations, model training, or complex setup procedures.

Joavku is designed to operate as a standalone component within a modular, real-time football analysis pipeline. It supports configurable preprocessing steps such as gamma correction and dynamic kit cropping, and can adapt to new team kits without retraining. The system’s performance is evaluated through a series of experiments showing that it delivers reliable results at speeds suitable for live analysis. This work demonstrates how focusing on a specialized subtask, team assignment, can lead to improved efficiency, flexibility, and accuracy when compared to monolithic, end-to-end systems. Joavku contributes to the growing field of accessible and real-time sports analytics by offering a practical alternative to model-dependent solutions.