Surya Namaskar Pose Classification Using MobileNetV2 with Transfer Learning and Data Augmentation

Abstract

In this paper, we propose a robust deep learning-based framework that efficiently enables the automated classification of six basic yoga poses, comprising the Surya Namaskar postures. To tackle the problem of human postures recognition in visual data, it comes with a transfer learning system with MobileNetV2 as the main architecture. The dataset used in this work was created by extracting all images of a specific pose from the scale dataset and was divided into training, validation, and test sets in a 80-10-10 ratio. In reducing overfitting and improving model generalization, extensive data augmentation techniques such as rotation, shifting, zooming and horizontal flipping were utilized. Thus, the model is trained in 2 steps, first the base model is frozen and then the last 30 layers are unfrozen and fine-tuned. CP class weights were computed to cope with class imbalance. The system provides high classification performance in the test set so proving its efficiency. Performance was evaluated through confusion matrixes, classification report and ROC curves which assures the reliability and precision of the model. Our contribution to AI-assisted fitness and wellness lies in providing a resource-efficient, device-independent, accurate model for yoga pose recognition that can be used in areas such as personal training, rehabilitation, and remote wellness and fitness monitoring.