YAR-Net: Yoga Asanas Recognition Network Based on Variational Autoencoder

Yoga is a holistic practice aimed at rejuvenating individuals physically, mentally, and spiritually. Despite rapid technological advancements, there remains ample room for computational exploration across various social domains. Yoga image analysis is a nascent field that aims at finding the exact location of yoga postures so that they can recognize the yoga asana more reliably. Yet, yoga asanas recognition is a challenging task that is worsened due to increased dimensionality, and the problem of segmentation, leading to difficulties in recognizing asanas with high precision. To address these challenges, a variational autoencoder-based Yoga asanas recognition network (YAR - Net) is proposed, capable of reliably identifying numerous yoga asanas. The proposed methodology incorporates segmentation from Mask Region-based Convolutional Neural Network (mR-CNN) to identify the Region of Interest (ROI) followed by feature extraction from segmented ROI utilizing Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), and Edge features. Finally, recognition of yoga asanas is accomplished by YAR - Net which provides the recognition accuracy of 97.46%, and 99.91% using benchmark datasets Yoga107 and YogaPoses, respectively. The results underscore the YAR - Net's remarkable performance as compared to the state-of-the-art techniques.