Multi-Classification for Yoga Pose Based on Deep Learning
Abstract
The pose estimation technique has been widely utilized in physical therapy and sports areas, but this technique is barely used in the yoga domain. Yoga pose classification is relatively difficult when using neural networks because poses in yoga are more complex than normal poses. Therefore, this paper proposed a new method that combines post estimation algorithm and Convolutional Neural Network (CNN) to classify yoga poses. The pose estimation algorithm used in this paper is OpenPose algorithm. This algorithm is applied to detect the skeleton of a person and draw it onto a pure black picture. After extracting poses from original imagines, a Convolutional Neural Network is used to classify yoga poses and reaches 92.99% validation accuracy. This paper also compares the performance of models with and without OpenPose algorithm and the result shows that the accuracy of models without the assistance of OpenPose algorithm averages 3%- 6% lesser than the accuracy of models combined with OpenPose algorithm. Therefore, the proposed scheme of classification was proved to be effective based on the skeleton information extracted by OpenPose.