Viewpoint-Agnostic Taekwondo Action Recognition Using Synthesized Two-Dimensional Skeletal Datasets

Area of Science:

• Computer Science
• Sports Science
• Artificial Intelligence

Background:

• Taekwondo poomsae evaluation lacks objectivity due to the absence of standardized methods.
• Inconsistent scoring and fairness issues arise from subjective judging in Taekwondo poomsae.
• Current evaluation methods struggle with viewpoint variations, impacting accuracy.

Purpose of the Study:

• To develop an objective evaluation system for Taekwondo poomsae using artificial intelligence.
• To enhance the consistency and fairness of Taekwondo poomsae scoring.
• To create a robust action recognition model for Taekwondo poomsae adaptable to various viewpoints.

Main Methods:

• Utilized a three-dimensional (3D) convolutional neural network (CNN) for action recognition.
• Employed a full-body motion-capture suit to collect 3D skeleton data of Taekwondo poomsae.
• Generated synthesized 2D skeletons from multiple viewpoints to create a diverse training dataset.

Main Results:

• The proposed 3D CNN model demonstrated superior performance in recognizing Taekwondo poomsae actions.
• The model achieved robust and consistent recognition irrespective of camera viewpoints.
• Performance evaluation against 2D skeletons and RGB images confirmed the model's effectiveness.

Conclusions:

• The 3D CNN-based model offers an objective and reliable method for Taekwondo poomsae evaluation.
• This AI-driven approach can significantly improve fairness and consistency in Taekwondo competitions.
• The model's viewpoint-invariant recognition capabilities address a key challenge in automated sports analysis.