In artistic gymnastics, performance analysis tools that provide immediate feedback to athletes and coaches are key to optimizing performance. Therefore, we developed a system that provides direct video feedback and automatically identifies key frames relevant to achieve high performance. This system estimates the gymnast`s two-dimensional (2D) and three-dimensional pose from a single-view video and uses a long short-term memory (LSTM) network to identify key frames, in which contact with the ground or an apparatus starts or ends. We compared three 2D pose estimation algorithms and found that MoveNet yielded the highest accuracy, being able to detect 84% of the joints accurately. The LSTM network detected the key frames with a 97% F1 score. In conclusion, our system provides direct video feedback to gymnasts and coaches relevant performance parameters.
© Copyright 2024 ISBS Proceedings Archive: Vol. 42: Iss. 1. Alle Rechte vorbehalten.