Blaze Pose Graph Neural Networks and Long Short-Term Memory for Yoga Posture Recognition

Abstract

Yoga posture recognition plays a vital role in guiding practitioners and preventing injuries. This research proposes a deep learning-based system that integrates Blaze Pose, Graph Neural Networks (GNN), and Long Short-Term Memory (LSTM) to improve the precision of yoga posture classification. The system begins by extracting 33 key points of the human body from input images using Blaze Pose, a lightweight and efficient pose estimation model optimized for real-time applications. These key points are then modeled into a graph structure and processed using GNN to capture spatial dependencies among joints. To handle the temporal dynamics of movement, an LSTM network is employed, enabling the system to recognize postures across sequential frames. The dataset used comprises 2,139 images categorized into eight yoga poses, including static and transitional movements. Preprocessing steps include image resizing, normalization, and key point extraction. Evaluation metrics, such as precision, recall, and F1-score, indicate strong performance across all classes, with particularly high scores for poses like "tree" and "triangle". Despite some misclassifications between similar poses like "cobra" and "warrior", the overall confusion matrix shows high predictive reliability. Experimental results demonstrate that the proposed GNN and LSTM model achieves a training accuracy of 99.70% and a test accuracy of 97.81%, outperforming models based solely on LSTM or GNN. The findings confirm that combining GNN and LSTM enhances the model's ability to learn both spatial and temporal patterns in human pose data. This system demonstrates potential as an intelligent yoga assistant, capable of delivering real-time feedback with high accuracy.