Authors:
Rikuhei Umemoto
1
;
Kazushi Tsutsui
2
and
Keisuke Fujii
1
Affiliations:
1
Graduate School of Informatics, Nagoya University, Japan
;
2
Graduate School of Arts and Sciences, The University of Tokyo, Japan
Keyword(s):
Machine Learning, Soccer, Spatiotemporal Data, Prediction.
Abstract:
Analyzing team defense in soccer is challenging due to limited labeled data. Some previous methods for evaluating soccer defenses were based on the prediction of defensive events using the locations of all players and the ball. However, they did not consider the importance of multiple events and assumed perfect observation of all 22 players, which is not open-source, with a larger amount for learning the classifier. In this paper, we propose a generalized valuation method for defensive teams by score-scaling the predicted probabilities of events, including gaining possession of the ball and being attacked. Our method can be applied to the open-source location data of all players in frames from broadcast video of events, such as football games from Euro 2020, by investigating the effect of the number of players on event prediction performance. Our validation results using Euro 2020 data show that event prediction accuracy can be maintained with a limited number of player features for
scoring, conceding, gaining the ball, and effective attacks. Additionally, our defensive metric effectively explains the defensive characteristics and strengths of the top four teams in the tournament, while also highlighting the reasons why some teams received poor defensive evaluations. Our approach offers a practical way to analyze and evaluate team defenses even with self-recorded or broadcast videos.
(More)