Yoga pose recognition using dual structure convolutional neural network.

As a popular form of physical and mental exercise, the correct execution of yoga movements is crucial. With the development of deep learning technologies, automatic recognition of yoga postures has become popular. To recognize five different yoga postures, this article proposed a dual structure convolutional neural network with a feature fusion function, which consists of the convolutional neural network A (CNN A) and convolutional neural network B (CNN B). Among them, the structure CNN A observes different channels finding the global feature of yoga images, and the structure CNN B calculates the depth information in each pixel of the yoga images. Following that, the extracted global feature and local feature are fused by a feature fusion function of taking a matrix dot multiplication. Finally, the softmax layer accurately recognizes yoga postures based on the fused features. Experimental results show that the proposed model achieves 97.23% accuracy with 96.08% precision and defeats against the competitors in the recognition of yoga postures. Moreover, the feature fusion function is proved to be successful in terms of the recognition to yoga postures. We also find that the feature fusion with a matrix dot multiplication operation can significantly improve the recognition accuracy of yoga postures than that with a direct connection operation.