Yoga Vision: Yoga Pose Detection and Correction System Using CNN
Jayesh Mohanrao Sarwade, Prachi Kulkarni, Siddhi Bhabad, Abhilasha R. Patil, Shraddha Choudhari · 2024
Yoga, originating from India, has garnered significant global traction owing to its manifold benefits for mental and physical well-being. Recent statistics reveal a staggering count of 300 million practitioners worldwide, with a growing number of instructors annually. Nonetheless, the prevalence of incorrect yoga postures underscores the critical necessity for precise alignment and a robust mechanism for identifying and rectifying deviations. This paper introduces a novel system for real-time yoga posture detection and correction leveraging MediaPipe for computer vision. Utilizing a fusion of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM), our deep learning architecture facilitates the recognition of yoga poses in live video streams. The CNN component is tasked with feature extraction from key body points, while the subsequent LSTM layer deciphers temporal dependencies within the sequence of frames for accurate pose predictions. Poses are classified as either correct or incorrect, and in the event of a correct pose, the system delivers pertinent feedback through text or speech modalities. Functionally, our system captures live footage of the practitioner via camera feed, extracts essential pose features, and matches them against a trained data model comprising a repertoire of known yoga poses. Upon detecting an incorrect posture, real-time corrective feedback is promptly relayed to the practitioner.