YOGA POSTURE DETECTION

Abstract: Yoga posture correction and recognition systems can help learners practice safely and consistently by providing instant feedback without requiring continuous supervision from an instructor. This project proposes a deep learning–based Yoga Posture Detection system that identifies yoga poses from images and real-time webcam video. The system uses a convolutional neural network (CNN) based classifier trained on a labeled dataset of yoga postures, where each class corresponds to a specific asana. For real-time operation, YOLOv8 is used to detect the person in each frame, the detected region is cropped, and the posture is classified using the trained model. The classification model is trained using transfer learning (MobileNetV2 backbone) to improve accuracy with a limited dataset and reduce training time. The final system is deployed as a web application using Flask with a user-friendly interface built using HTML, CSS, and JavaScript, allowing users to upload images for posture prediction and view top confidence results. Experimental results show that the proposed model achieves around 70% validation accuracy over 41 yoga classes, and performance is analyzed using a confusion matrix, classification report, and Grad-CAM visual explanations. The solution demonstrates an end-to-end pipeline for yoga pose classification and real-time detection, and can be extended further for posture correction and fitness guidance.

Keywords: deep learning, yoga pose detection, Convolutional Neural Network (CNN), Real-Time Detection.