Transfer Learning with Deep Representations is Used to Recognition Yoga Postures

Abstract

Human activity identification is the automated interpretation of the movements happen in a video done by a human. Iterative Due to its wide applications in fields such as autonomous driving, biomedical imaging, and machine intelligence vision, among others, recognizing human activity in an image remains a tough and crucial research subj ect in the field of computer vision. Deep learning techniques have recently advanced, and models for image identification and classification, object detection, and speech recognition have been successfully implemented. Only a few examples include different aspects of human structure and movement, diffraction, a busy background, and so on. Moving cameras, changing lighting conditions and changing perspectives are all things to think about. Yoga is an excellent kind of physical activity. It's critical to maintain proper yoga posture. This research provides a unique technique for yoga asana detection based on feature extraction and representation Using a deep CNN model that has already been trained, followed by yoga asana recognition using a hybrid Support Vector Machine (SVM) and K-Nearest Neighbor (KNN) classifier. With the constrained training datasets, it was discovered that previously learned CNN-based representations on large-scale annotated datasets may be applied to yoga asana recognition tasks. In real-time datasets, the suggested approach is tested on seven yoga asana (Pranamasana, Dhanurasan, Dandasana, Gomukhasan, Garudasana, Padmavrikshasana and Padmasan). The results show that the proposed scheme outperforms the state of the art methods.