Improvement of accuracy of pose estimation of soccer shooting action from movies

Abstract

In recent years, there has been a demand for a platform in amateur soccer that can easily analyze the performance of a game from player data. When accumulating and analyzing player data, the platform needs a function that recognizes the player's shooting motion. However, when posture estimation is performed for a video in which a player is shooting, there is a problem that key points of both hands and both feet immediately after kicking the ball may be undetected or erroneously detected. In this paper, we are working on improving the accuracy of the posture estimation of a player during a shooting motion. Specifically, after recruiting subjects and collecting data, the part where key points are not detected / misdetected is corrected, a teacher label is created, machine learning is performed, and the detection rate is improved.