Objective Detection of High-Risk Tackle in Rugby by Combination of Pose Estimation and Machine Learning

Monami Nishio, Naoki Nonaka, Ryo Fujihira, Hidetaka Murakami, Takuya Tajima, Mutsuo Yamada, Akira Maeda, Jun Seita

Abstract: To provide suitable care for concussion, objective and timely detection of high-risk event is crucial. Currently it depends on monitoring by medical doctors, and there is a certain risk of missing high-risk events. A few attempts introducing video analysis have been reported, but those approaches require labeling by experts, which is skill-dependent, and time and cost consuming. To achieve objective detection of high-risk tackle without human intervention, we developed a method combining pose estimation by deep learning and pose evaluation by machine learning. From match videos of Japan Rugby Top League in 2016–2018 seasons, 238 low-risk tackles and 155 high-risk tackles were extracted. Poses of tackler and ball carrier were estimated by deep learning, then were evaluated by machine learning. The proposed method resulted AUC 0.85, and outperformed the previously reported rule-based method. Also, the features extracted by the machine learning model, such as upright positions of tackler/ball carrier, tackler’s arm dropped in extended position, were consistent with the known risk factors of concussion. This result indicates that our approach combining deep learning and machine learning opens the way for objective and timely detection of high-risk events in rugby and other contact sports.