Yoga Pose Estimation Using MoveNet Preprocessor and Deep Learning Technique

This study explores yoga pose estimation using the MoveNet model, a deep learning framework, to extract key points from images. We hypothesize that MoveNet can accurately identify and analyze key points in various yoga poses. Our methodology involves implementing TensorFlow and Keras-based algorithms to process a diverse dataset of approximately 8300 yoga posture images. Utilizing the Yoga-82 dataset, which formulates pose estimation as a classification task across 82 different classes, we conduct data preprocessing to ensure compatibility with MoveNet. Key point extraction is performed using MoveNet’s bottom-up pose estimation approach, which employs a heatmap-based technique for precise localization of human keypoints. Classification is then carried out using an Artificial Neural Network (ANN). While the training accuracy exceeds 90%, the validation accuracy remains under 90% for the overall dataset. Notably, certain poses exhibit accuracy rates exceeding 99% despite minimal representation. This research showcases the potential of deep learning in enhancing physical training and well-being, emphasizing the importance of robust data handling methodologies in yoga pose estimation tasks. Through a comprehensive exploration of hypothesis, methodology, techniques, and tools utilized, this study contributes to the intersection of health, fitness technology, and AI, paving the way for personalized health and fitness initiatives.