Abstract: This study suggests a thorough deep learning framework for identifying yoga poses and providing in-the-moment instructions. Using a standard RGB webcam, a custom dataset comprising six commonly performed asanas—Bhujangasana, Padmasana, Shavasana, Tadasana, Trikonasana, and Vrikshasana—was captured indoors Using a hybrid CNN–LSTM architecture and Open Pose-based pose estimation, the method models temporalcontinuity and spatial key point configurations in brief sequences. Configurations andtemporal continuity in short sequences. Three tests Three settings are temporal voting on45-frame windows (~15 s), frame-wise classification, and real-time webcam inference withan invisible participant. In real-time evaluations, the system attains 9892% accuracy,9938% accuracy at the windowed vote level, and 9904% accuracy at the frame level. (i) asuccinct yet effective spatiotemporal model for yoga recognition; (ii) a reproduciblepipeline made entirely of RGB inputs; (iii) the elimination of temporal pooling and thresholding strategies; and (iv) a publicly available data set complete with evaluation protocolsThe proposed framework offers a workable way to incorporate posture-awareness features into in-home coaching programs, rehabilitation settings, and consumer fitness applications.

Keywords: Human activity recognition, Yoga, Open Pose, CNN–LSTM, Spatiotemporal modeling, Real-time systems.

Downloads: | DOI: 10.17148/IARJSET.2025.12931