Yoga Pose Detection Using Posenet and k-NN

Abstract

Yoga offers a wide range of asanas, and the angle between body parts plays an important role here. This project carries a non-profit system that strives to develop core muscles using yoga-like poses. While practicing yoga asanas virtually, the proposed technique perfectly detects the human position. To contemplate the dissension of the angle formed with original values, the cosine similarity technique is applied. Multiple dimensions must be addressed since crucial angles are made up of a critical combination of angles. This system detects the difference between the actual and target positions and corrects the user by delivering real-time image output and necessary instructions to correct the identified pose. Human poses estimation is utilized in this research to estimate an individual's Yoga position using computer vision techniques and Open pose (open-source library). In most circumstances, the suggested method retains high accuracy while achieving real-time speed. The proposed model was trained with 90% of data and tested with 10% of same with real-time testing, resulting 94 % of accuracy.