TUHAD: Taekwondo Unit Technique Human Action Dataset with Key Frame-Based CNN Action Recognition

Area of Science:

• Sports Science
• Computer Vision
• Artificial Intelligence

Background:

• Taekwondo poomsae (form) competitions lack objective scoring, unlike Olympic sparring (gyéorugi), leading to fairness concerns.
• The need for quantitative evaluation tools in taekwondo poomsae is growing.
• Applying action recognition to taekwondo is challenging due to the extreme and rapid movements involved.

Purpose of the Study:

• To develop a quantitative evaluation tool for taekwondo poomsae using action recognition.
• To establish a multimodal dataset for taekwondo unit technique action recognition.
• To design and validate a convolutional neural network for recognizing taekwondo actions.

Main Methods:

• Creation of the Taekwondo Unit technique Human Action Dataset (TUHAD) with 1936 multimodal action samples from 10 experts across 8 unit techniques and 2 camera views.
• Development of a key frame-based convolutional neural network architecture for action recognition.
• Validation of the model's accuracy across various input configurations using correlation analysis.

Main Results:

• The proposed convolutional neural network model achieved a maximum recognition accuracy of 95.833% for taekwondo actions.
• The lowest accuracy recorded was 74.49%, indicating performance variability based on input configuration.
• Correlation analysis provided insights into how input configurations affect recognition accuracy.

Conclusions:

• The study successfully developed a taekwondo action recognition model and dataset (TUHAD).
• The findings contribute to the advancement of objective evaluation methods in taekwondo poomsae.
• This research paves the way for more equitable and accurate assessments in taekwondo competitions.