Conference Details

Abstract :

Exerting yoga asanas accurately without a trainer can be challenging. Computer vision techniques, such as pose detection and classification, have the potential to provide feedback on pose accuracy. This paper reviews current methodologies, identifies research gaps and explores opportunities for improvement in yoga pose detection and classification. An experimental transfer learning model is proposed that can be deployed on mobile devices. The model extracts features from images of accurately executed poses and uses them as input for machine learning and neural network models, enabling estimation of pose correctness. Handheld devices are more suitable for yoga pose detection and classification compared to stationary devices such as PCs, as they are portable and convenient for real-time feedback and adjustments. However, running such a model on mobile devices without optimization can be inefficient due to the high memory and computing requirements. This paper proposes the use of quantization and pruning to reduce memory access costs and increase compute efficiency, making the model more efficient for edge devices. Despite variability in individual performance, pose detection and classification through mobile devices could provide helpful feedback for yoga practitioners. Our optimized model occupies 15% of the traditional model’s size with an average error of 1.2%.