Yoga Asanas Pose Detection using Feature Level Fusion with Deep Learning-Based Model

Yoga is a healthful activity that started in India and has spread throughout the world, rejuvenating a man's physical, mental, and spiritual well-being. However, Yoga Image analysis is recent research based on Position identification. So Pose discovery procedures can recognize yoga postures and help people practise Yoga more precisely. But increasing dimensionality and segmentation problems lead to posture recognition is a complex problem. To resolve this, we propose a Deep learning algorithm that can be used to reliably identify numerous yoga asanas, as demonstrated in this paper. Since all the video data is raw, pre-processing is necessary before the model can be used. Objects are recognized using the YOLOv5 network in the pre-processed frames, and skeleton pose estimation is performed. Based on the identified objects, an algorithm makes a skeleton of a human in real-time. The skeletal position skeleton and objects detected by YOLOv5 are fused using feature-level fusion to derive the angles of the linkages in the human body. Based on Generative Adversarial Network (GAN) is implemented to predict the yoga postures, and these outputs are fed into the learning process. About 80% of the dataset was for training, while the residual 20% was used for testing purposes. The proposed system achieves high performance in identifying the Yoga Pose gestures accurately compared to other systems.