Transformer-Based Multi-Object Tracking of Football Players Using Pseudo-Labeling

Abstract

This thesis investigates the problem of tracking football players in video sequences, with a focus on adapting modern multi-object tracking (MOT) methods to the specific challenges of football environments. The work is part of a broader effort to develop tools for analyzing football games using automated visual data. In this context, we utilize MOTRv2, a transformer-based tracking model originally designed for general-purpose MOT tasks, and apply it to the football domain, where challenges such as frequent occlusions, tight formations, and rapid movement are prevalent. To address the lack of annotated football-specific tracking data, we implement a pseudo-labeling framework that allows the model to be trained on unlabelled domain data in a semi-supervised fashion. This approach enables progressive refinement of the model through multiple training cycles on domain-specific content. Our results show that MOTRv2 can be adapted to the football setting and performs well in many scenarios, particularly in open-play segments with clear player separation. However, limitations remain, including decreased tracking stability in crowded scenes and occasional ID-switches due to overlapping motion patterns.

Overall, this work demonstrates the potential of transformer-based trackers in sports applications and highlights the benefits of self-supervised training when domainspecific data is scarce. The findings offer insights for future improvements in automated sports tracking systems.