Background: The evaluation of Human Movements based on Taekwondo poomsae (movement patterns) is inherently subjective, leading to concerns about bias and inconsistency in scoring. This emphasizes the need for objective and reliable scoring systems leveraging artificial intelligence technologies. Objectives: This study seeks to enhance the accuracy and fairness of Taekwondo poomsae scoring through the application of camera-based pose estimation and advanced neural network models. Design: A comparative analysis was conducted to evaluate the performance of machine learning models on a large-scale Taekwondo poomsae dataset. Methods: The analysis utilized a dataset comprising 902,306 labeled frames from 48 participants performing 62 distinct poomsae movements. Five models—LSTM, GRU, Simple RNN, Random Forest, and XGBoost—were evaluated using performance metrics, including accuracy, precision, recall, F1- score, and log loss. Results: The LSTM model outperformed all others, achieving an accuracy, precision, recall, and F1-score of 0.81, alongside the lowest log loss value of 0.55. The GRU model demonstrated comparable performance, while traditional models such as Random Forest and XGBoost were less effective in capturing the temporal and spatial patterns of poomsae movements. Conclusion: The LSTM model exhibited superior capability in modeling the temporal and spatial complexities inherent in Taekwondo poomsae, establishing a robust foundation for the development of objective, scalable, and reliable poomsae evaluation systems.