Real-time yoga pose classification with 3-D pose estimation model with LSTM

Yoga has now become a part of life for humans all over the globe enabling them to unite the body with the mind, helping in achieving a healthy lifestyle in the modern day. Practicing yoga with a trainer is always recommended to get maximum benefit but with this fast-paced lifestyle getting time for a lesson with a trainer or under proper guidance gets difficult and also sometimes middle-class families cannot bear the cost of a trainer. Therefore, a system is required which is accessible to everyone and can help in performing yoga poses and improving. This paper introduces a novel framework developed for estimating yoga poses through computer vision using a 3-D top-down semantic key landmark estimator with a Recurrent Neural Network (RNN) for classification. For training and validation of our model, we tailored our custom dataset of 10 different yoga poses having a total of 300 sequences. The model on the dataset gave an average of 92.34% accuracy using Long Short-Term Memory (LSTM) classifier.