Yoga injury risk prediction model based on generative adversarial network

Incorrect execution of yoga postures, especially in unsupervised environments, increases the risk of musculoskeletal injuries. To address this challenge, a Yoga Injury Risk Prediction Model based on a Generative Adversarial Network (GAN) is proposed, named YOGA-GAN. This model aims to predict injury risks associated with incorrect posture execution. YOGA-GAN leverages the generative capabilities of GANs to create synthetic samples of rare or injury-prone yoga postures, enriching the dataset and mitigating class imbalance issues. The dataset consists of annotated images of both correct and incorrect poses. Preprocessing techniques, such as normalization and Gaussian filtering, are applied to enhance data quality. Feature extraction methods like Histogram of Oriented Gradients (HOG) and Scale-Invariant Feature Transform (SIFT) capture critical structural and textural features of each pose. Principal Component Analysis (PCA) is used for dimensionality reduction while retaining essential information. The discriminator in the GAN is adapted to serve a dual role: distinguishing between real and synthetic poses and predicting the injury risk level of each posture. The system is implemented using Python, providing a robust foundation for developing real-time yoga assistance systems aimed at reducing injury risks. Experimental evaluation demonstrates significant improvements in standard performance metrics, achieving accuracy ( 96.4% ), precision ( 97.1% ), recall ( 98.2% ), and F1-score ( 97.3% ) compared to traditional deep learning models. The YOGA-GAN framework presents a step forward in safer yoga practice and injury prevention.