Application of Human Posture Recognition Based on the Convolutional Neural Network in Physical Training Guidance

Area of Science:

• Sports Science
• Computer Vision
• Biomechanical Analysis

Background:

• Traditional sports human body posture estimation methods exhibit significant errors in complex environments, especially with occluded body parts.
• Accurate athlete pose estimation is crucial for effective training and competition analysis feedback.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) model for improved athlete pose estimation in sports game videos.
• To address the limitations of traditional methods, particularly in complex and occluded scenarios.

Main Methods:

• Development of a superimposed hourglass network integrating improved, multiscale, and large perception models.
• Implementation of intermediate supervision to mitigate the gradient disappearance problem in CNNs.

Main Results:

• The proposed CNN model significantly improves the accuracy of athlete pose estimation.
• The model effectively reduces the negative impact of occlusion on pose estimation accuracy.
• Demonstrated competitive advantages and high accuracy compared to existing methods under standard conditions.

Conclusions:

• The CNN-based athlete pose estimation model offers enhanced accuracy and robustness in sports video analysis.
• This approach provides a valuable tool for athlete training, competition feedback, and performance analysis, even in challenging conditions.