Learning Shape Priors: A Yoga Posture Recognition

This paper presents an automated model for real-time yoga posture recognition. The model consists of four steps: data collection, pose recognition using Open Pose, feature extraction with LSTM and CNN, and training with a polling approach. A data set of diverse yoga posture videos was collected and divided into training, testing, and validation sets. We used the Open Pose library for accurate pose extraction from video frames. The model combines CNN and LSTM to extract features from key-points and analyze temporal changes. A Soft-max layer calculates the probability of each yoga step, enabling precise posture identification. Real-time feedback is provided by aggregating predictions over a 1.5second time window using polling. The experimental evaluation demonstrated the effectiveness of the approach. Frame-by-frame analysis achieved high accuracy of 99.04 %. To enhance real-time predictions, polling improved overall accuracy. The developed model offers real-time feedback for individuals practicing yoga, aiding in posture correctness and enhancing the yoga experience. This innovation has the potential to transform the field of yoga practice.