Abstract:
"Traditional yoga teaching techniques require a physical presence, which often constrains the
scalability and accessibility of training resources for learners. Furthermore, the absence of real-
time, objective feedback makes self-practice difficult and potentially hazardous due to
improper poses. Current digital yoga platforms fall short in providing accurate, immediate, and
personalised feedback, necessitating the need for a technologically advanced solution that is
scalable, accessible, and accurate.
This work introduces an innovative method to address the previously mentioned issue. The
core of the solution lies in a model utilising a Spatial Graph Convolutional Neural Network
(SGCNN) to analyse real-time yoga postures. The advantage of deep learning techniques is
harnessed along with skeletal data captured from a 3D camera to facilitate precise pose
estimation and correctness assessment. The SGCNN model is educated through a rich dataset
that contains a wide range of yoga postures demonstrated by individuals of various body types
and skill levels.
The model was tested for pose recognition accuracy and real-time interaction. SGCNN's
recognition accuracy was 87%, surpassing standard machine learning methods. The real-time
feedback system performed well with an average response time of 0.3 seconds, ensuring
smooth engagement. A user assessment of yoga practitioners of various skill levels confirmed
the concept, with most saying the system helped their practises by delivering correct feedback.
This supports the idea that the model would revolutionise yoga and digital fitness by making it
safer and more accessible."
