[3N5-GS-7] Vision, speech media processing:

[3N5-GS-7-01] Action Prediction using Goalkeeper's Joint Coordinates in Soccer Penalty Kicks

〇Keito Abe¹, Haruki Tsubouchi¹, Koushi Matsuzawa¹, Zaiying Zhao¹, Tomoya Sugihara¹, Toshihiko Yamasaki¹ (1. Univ. of Tokyo)

Keywords:Object Detection, Motion Prediction, Transformer

In penalty kick (PK) situations in soccer, previous studies have focused on using machine learning to predict the direction of the kicker's shot, but studies on goalkeeper’s movement are limited. To address this, we propose action prediction models of the goalkeeper, which is useful for the PK kicker. In this study, we develop two models to predict goalkeeper movements during PK scenarios and evaluate their performance. Using You Only Look Once (YOLO), we make standardized joints-coordinates dataset from various kinds of soccer videos. The input of the models is time-series data of goalkeepers' joint coordinates. As an output we consider two problems: one is a regression task of the time-series data of the goalkeeper's joint coordinates after the kick, and the other is a classification task to predict the region where the goalkeeper will pass. We use LSTM and Transformer models and the Mean Per Joint Position Error (MPJPE) for the regression task is down to 1260mm and the accuracy for the classification task is 0.71. Our study shows the feasibility of predicting goalkeeper movements and potentially be useful for the kicker in PK situations.