Abstract:
"Taekwondo, a martial art that originated in Korea in the mid-20th century, has now evolved into one of the
most widely practiced martial arts globally. Within Taekwondo, there are two main types of competitions
namely poomsae and gyeorugi. While gyeorugi involves controlled combat between two competitors aiming
to score points, poomsae competitions focus on executing patterns of Taekwondo techniques, such as stances
and strikes, and are judged based on the accuracy and power of the techniques.
Stances are the basis upon which all other techniques are built on. It affects precision, efficiency, and stability
of techniques. Therefore, it is important for beginners as well as advanced practitioners to master correct
stances as a primary focus on their training.
This research proposes a system designed to serve as a training environment for independent learning of
Taekwondo martial art by classifying and correcting Taekwondo stances in real-time.
Considering the specific domain of Taekwondo and the constraints made by limited data, this research
employs XGBoost classification algorithm with MediaPipe, achieving an impressive accuracy of 94.8%.
Additionally, a few pretrained CNN models are employed with XGBoost for classification to conduct
comparative analysis. This research aims to contribute to both the domains of pose estimation and Taekwondo
by demonstrating the practical application of pose estimation techniques within martial arts."
