Pose Fusion AI Framework for Personalized Yoga Asana Guidance Alignment Analysis and Performance Assessment

Abstract

This paper presents a real-time AI yoga pose instructor that combines person detection, landmark-based geometric representation, temporal deep learning, referencepose comparison, and web-based feedback. The implemented pipeline detects the practitioner with YOLO11, extracts 33 body landmarks through MediaPipe, converts each valid frame into a 56-dimensional vector containing normalized coordinates, body distances, and joint angles, and forms 30-frame sequences for pose classification. Four temporal and convolutional models— LSTM, GRU, one-dimensional CNN, and ConvLSTM—were trained and evaluated on 43 videos of Tadasana, Trikonasana, and Vrikshasana with video-level train/validation/test separation. The GRU achieved the strongest test performance with 99.31% accuracy, 99.32% precision, 99.31% recall, and 99.31% weighted F1-score on 433 test sequences. Beyond pose identity, the deployed Flask/SocketIO application compares the live pose with reference landmarks, computes landmark and angle similarity, identifies incorrect joints, and provides corrective feedback. A real-time deployment evaluation demonstrates that high pose-recognition confidence can coexist with moderate posture alignment; for example, Tadasana was recognized with 98.66% GRU confidence while posture accuracy was 82.61%, producing a “Needs Improvement” status. The system therefore separates pose recognition from posturequality assessment and provides interpretable real-time guidance.