Yoga pose recognition using dual structure convolutional neural network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Sports Science

Background:

• Yoga's popularity as physical and mental exercise necessitates accurate movement execution.
• Deep learning advancements have spurred interest in automatic yoga posture recognition.

Purpose of the Study:

• To propose a dual structure convolutional neural network (CNN) with a feature fusion function for recognizing five different yoga postures.
• To evaluate the effectiveness of feature fusion, specifically matrix dot multiplication, in enhancing recognition accuracy.

Main Methods:

• A dual CNN architecture (CNN A and CNN B) was developed.
• CNN A extracts global features from different image channels.
• CNN B calculates pixel-wise depth information, and features are fused using matrix dot multiplication before softmax classification.

Main Results:

• The proposed model achieved 97.23% accuracy and 96.08% precision in yoga posture recognition.
• The feature fusion function proved successful in improving recognition performance.
• Matrix dot multiplication fusion significantly outperformed direct connection fusion.

Conclusions:

• The dual CNN model with matrix dot multiplication feature fusion is highly effective for automatic yoga posture recognition.
• This approach offers a significant improvement over existing methods.
• Accurate yoga pose recognition has implications for practice and training.