Yoga Pose Detection Using Long-Term Recurrent Convolutional Network

Yoga is an ancient art form originated in India that makes one fit physically and also provides mental peace/relaxation. Yoga has become popular all over the world because of the increase in stress in modern lifestyle. A model has has been developed which detects various yoga poses. An open-source dataset of 88 videos consisting of 6 asanas performed by 15 different volunteers have been used. Various approaches have been studied that are available for pose detection and the most suitable approach has been chosen to meet the requirements. A combo of Long Short Term Memory (LSTM) and Convolutional Neural Network (CNN), called Long-Term Recurrent Convolutional Network (LRCN) has been put to use. LSTM is used to make temporal prediction, whereas CNN is used to extract information from each frame. The proposed model differs from existing models as it does not use open pose or pose net for keypoint detection. Using the proposed model, accuracy of 81 achieved.