Semi-automated last touch detection for out-of-bounds possession decisions in football

Abstract

Football referees must make quick and accurate decisions in unforgiving environments. In parallel, advances in optical tracking have created new avenues for technology-assisted officiating. Using skeletal and ball tracking data, we present a novel diphase framework for Semi-automated Last Touch detection, designed to help referees adjudicate out-of-bounds possession decisions where player and ball occlusion may pose challenges. The proposed methodology uses a touch probability model to find the decision frame of the last touch before the ball goes out-of-bounds, and rules-based or supervised learning algorithms predict the player responsible for the touch. Leveraging principles of kinematics, human anthropometry, and machine learning, the models predict the correct possession decision with up to 82.5% accuracy on a test dataset of duels from the 2022 FIFA World Cup, including over 90% for aerial duels. Our results represent potential improvements in human performance reported in previous literature and provide a baseline benchmark for future studies.