Background: The growing need for objective and accurate evaluation in Taekwondo poomsae competitions has highlighted the limitations of subjective human judgment. Objectives: This study aims to develop an automated scoring framework using camera-based pose estimation and advanced neural networks to improve the consistency and accuracy of poomsae evaluation. Design: Comparative analysis of neural network architectures on a large-scale dataset of poomsae movements. Methods: A dataset of 902,306 labeled frames, captured from 48 participants performing 62 distinct movements using synchronized multi-view cameras, was analyzed. Five neural networks (HNN, 1D CNN, GCN, MLP, SANN) were implemented and evaluated using accuracy, precision, recall, and F1-score. Results: The HNN demonstrated superior performance with an F1-score of 0.78 in classifying Taekwondo poomsae postures. The 1D CNN followed with an F1-score of 0.76, while GCN, MLP, and SANN achieved F1-scores of 0.74, 0.70, and 0.66, respectively. The HNN's hierarchical feature extraction approach proved effective in capturing the complex spatial and temporal patterns inherent in poomsae movements. Conclusion: Hierarchical Neural Networks outperform other architectures in poomsae classification, establishing a foundation for objective and scalable scoring systems in competitive settings.

INTRODUCTION
METHODS
    Data Collection and Preparation
    Data Augmentation and Feature Engineering
    Model Implementation
    Training and Optimization
    Evaluation Metrics
    Experimental Workflow
RESULTS
DISCUSSION
CONFLICTS OF INTEREST
REFERENCES

• Taeseok Choi(Department of Physical Therapy, Gunjang University, Gunsan, Republic of Korea)
• Seoyoon Heo(Department of Occupational Therapy, Kyungbok University, Namyangju, Republic of Korea) Corresponding author