Asana AI – Yoga Pose Detection Using Pose Estimation and Real-Time Feedback System

The increasing global focus on health and fitness has led to a surge in yoga practitioners, many of whom seek digital assistance to improve pose accuracy and safety. This research, titled Asana AI: Real-Time Yoga Pose Detection and Correction System, leverages machine learning and computer vision to identify yoga poses and assess their correctness through real-time or uploaded video frames. The system employs Media Pipe for pose landmark extraction and a custom-trained classifier based on supervised learning for pose recognition. The application provides audio feedback— prompting corrections or affirmations—based on the detected pose and its accuracy. The system supports both webcam and image upload functionalities, making it accessible and user-friendly. The dataset consists of curated yoga videos and images from the Mendeley Yoga Dataset, focusing on four core asanas: Anantasana, Bhujang asana, Tadasana, and Vajrasana. Our approach achieved high training and validation accuracy of 90.62% and 90.17%, respectively, with 144,026 trainable parameters, demonstrating that accurate yoga pose correction which can be implemented in a computationally efficient and cost-effective manner. The application provides audio feedback—prompting corrections or affirmations—based on the detected pose and its correctness. It supports both webcam and image upload functionalities, making it accessible and user-friendly.