From Pixels to Perfect Form: Deep Learning for Real-time Yoga Pose Analysis
Preeti Garg, Karnika Dwivedi, Bharti Chugh, Madhu Gautam · International Journal on Smart & Sustainable Intelligent Computing · 2024
For maintaining a healthy life, yoga plays a very important role. Now, a day's people are using online platforms to learn yoga poses and maintain a healthy life. While doing yoga the most important point to take care of is maintaining the correct yoga posture. To learn how to recognize suitable yoga poses and give feedback to improve posture, this research presents a deep learning-based method for estimating yoga posture. The model is period-ically fed frames from videos or pictures. Key points are extracted with Keras multi-person pose estimation, yielding 12 joint vectors. The angles of these vectors relative to the x-axis are calculated. One of the six yoga positions is identified by applying a classification model based on these angles. A dataset with six distinct yoga poses, resulting in 70 videos and 350 instances, is used to test the methods. There are thirty validation examples, three hundred training instances, and thirty testing instances in the dataset. Various approaches such as Support Vector Machines (SVM), Convolutional Neural Networks (CNN), and CNN in conjunction with Long Short-Term Memory (CNN + LSTM) are implemented, and results reveals a competitive performance mosaic. Though CNN and CNN + LSTM architectures are believed to be superior, the updated feature set allows Multi-Layer Perceptron (MLP) to achieve an im-pressive accuracy of 0.9958