Yoga Pose prediction using Transfer Learning Based Neural Networks

Abstract

Abstract The rapid increase of the technology made human to work harder irrespective of the body condition. This has caused many people with mental pressure and lack of body fitness. Covid – 19 has also shown the serious effect on human mental health. Therefore, many physicians suggested to practice yoga for breathing and physical fitness. But it is not possible for all people to come to fitness centers in lockdown, tight work schedule, staying far from fitness centers etc. So, this research uses transfer learning approach on convolutional neural networks in order to predict the yoga posture made by the person in real – time. 85 different types of yoga postures dataset is constructed by web scrapping and by capturing the images which is more compared to existing works. Training, validation and test set were divided in the ration 70:10:20 respectively. The pre-trained models like EfficientNet – B0, Xception, ResNet – 50 and MobileNet were chosen based on their past performances and were trained on the constructed yoga dataset. The experimental results shows that, Xception model using transfer learning gave the best results with the testing accuracy of 95.67% and also second best in execution time.