Score prediction using multiple object tracking for analyzing movements in 2-vs-2 Handball

Ren Kobayashi, Rikuhei Umemoto, Kazuya Takeda, Keisuke Fujii

Abstract: Automatic evaluation of players’ movement is an important issue in sports. In handball, researchers have realized it using sensors of local positioning systems and machine learning. However, we cannot always use such sensors, especially in nonprofessional teams. Here, we propose a system to evaluate players’ movement using multiple object tracking from videos and logistic regression models trained with score labels. Using handball 2-vs-2 videos, we showed that the proposed system predicted a score with an average accuracy of over 60%. We also analyzed the factors contributing to the score prediction.