Using Machine Learning to Differentiate Set Pieces in Football via Tracking Data

Abstract

Event data- compiled time-stamped list of important events during football matchesare key sources of analysis for teams, analysts, and fans of the sport. A recent undertaking by the MIT Sports Lab and FIFA has resulted in the creation of an algorithm to generate event data from player tracking data, removing the need for manual compilation. The algorithm performs well, but possesses edge cases for which it cannot distinguish events due to algorithmic or data limitations. We propose and test a learning-based approach to classifying set pieces from tracking data, aiming to use differences in ball and/or player motion to inform us of which set piece corresponds to a dead-ball interval. The model shows promise in distinguishing corners, free kicks, and throw-ins. While far from ready to be utilized in a real game scenario, it shows the potential viability in distinguishing a certain class of events without relying on noisy ball data.