Spatial-Temporal Graph Convolutional Framework for Yoga Action Recognition and Grading

Abstract

The rapid development of the Internet has changed our lives. Many people gradually like online video yoga teaching. However, yoga beginners cannot master the standard yoga poses just by learning through videos, and high yoga poses can bring great damage or even disability to the body if they are not standard. To address this problem, we propose a yoga action recognition and grading system based on spatial-temporal graph convolutional neural network. Firstly, we capture yoga movement data using a depth camera. Then we label the yoga exercise videos frame by frame using long short-term memory network; then we extract the skeletal joint point features sequentially using graph convolution; then we arrange each video frame from spatial-temporal dimension and correlate the joint points in each frame and neighboring frames with spatial-temporal information to obtain the connection between joints. Finally, the identified yoga movements are predicted and graded. Experiment proves that our method can accurately identify and classify yoga poses; it also can identify whether yoga poses are standard or not and give feedback to yogis in time to prevent body damage caused by nonstandard poses.