Abstract Pose estimation of human activity recognition has been a keen area of interest in augmented reality experiences, gaming and robotics, animations, behavioral analysis, and more. One such exciting variant of pose estimation in the field of health and science is yoga pose estimation. This paper explores yoga pose estimation using deep learning networks. The research aims to build a system for estimating 45 different complex yoga asanas from 11,000 images using deep learning algorithms. This system is built using a Region-based Convolutional Neural Network (RCNN) to estimate the joints in the body, followed by a Convolutional Neural Network (CNN) for classifying the poses. The model is trained using the Yoga-82 (hierarchically labeled) dataset, a new dataset with complex pose variations mainly designed for hierarchical labeling. Next, it highlights the pose estimation task through ResNet models followed by an optimization algorithm, which increases the accuracy by 10%. The resultant accuracy is 90.5% for the ResNet50 model. Finally, it provides a solution for overlapping yoga poses, multi-person, in-air, and non-conventional poses using a dense network of 17 critical points for analysis and prediction.
Alan : Mühendislik
Dergi Türü : Uluslararası
Benzer Makaleler | Yazar | # |
---|
Makale | Yazar | # |
---|