Research on Different Convolutional Neural Networks in the Classification Scenes of Yoga Poses based on OpenPose Extraction

Abstract

The assessment of yoga pose is a very challenging problem, which did not receive much attention in the past, however, in fact, it has great use in daily exercise or medical rehabilitation. In this paper, the experimental study completed the extraction of skeleton information of yoga poses and classifies them with convolutional neural network. The experiment uses Openpose algorithm obtain the joint points of the original dataset image, then it connects the different joints together appropriately, to obtain the skeleton image. Subsequently, the research carried out the construction work of four different convolutional neural networks, among which VGG16, VGG19 and MobileNet three networks directly used the network model that has been built, and the self-built convolutional neural network set parameters layer by layer, and completed the network construction. After that, the research used the original image dataset and the skeleton image dataset train the four convolutional neural networks, respectively, and compared the accuracy of the test set, thus analyzing the impact of the two datasets on the training of the network model and the classification performance of each network. The experimental results show that the skeleton image dataset has better performance in network model training in comparison with the original image dataset, thus proving the availability of the method proposed in the research.