Sit-and-Reach Pose Detection Based on Self-Train Method and Ghost-ST-GCN.

Area of Science:

• Sports Science
• Biomechanical Analysis
• Artificial Intelligence in Physical Education

Background:

• The sit-and-reach test is crucial for adolescent flexibility and is mandatory in Chinese physical fitness assessments.
• Incorrect sit-and-reach postures can lead to sports injuries like muscle strains.
• There is a need for accurate methods to evaluate sit-and-reach technique during practice.

Purpose of the Study:

• To propose and evaluate an efficient AI model for judging the correctness of the sit-and-reach pose.
• To improve the accuracy and efficiency of pose estimation for physical exercises.
• To provide real-time feedback for adolescents practicing the sit-and-reach exercise.

Main Methods:

• A Ghost-ST-GCN model was developed for sit-and-reach pose correctness judgment.
• Seven body keypoints were detected using a self-training method with the BlazePose network.
• The model incorporates ghost layers within GCN-TCN blocks for enhanced efficiency.

Main Results:

• The self-training method improved keypoint annotation accuracy.
• Ghost layers streamlined the pose detection model.
• The system achieved 85.20% action detection accuracy for sit-and-reach with <1s latency.

Conclusions:

• The Ghost-ST-GCN model effectively and efficiently assesses sit-and-reach exercise correctness.
• This AI-driven approach can guide adolescents in standardizing movements and preventing injuries.
• The technology is suitable for independent practice and physical fitness evaluations.