Temporal Pattern Attention for Multivariate Time Series of Tennis Strokes Classification

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning
• Image Processing
• Human Action Recognition

Background:

• Human Action Recognition (HAR) is crucial for understanding human behavior in applications like sports analysis.
• HAR aids in evaluating player performance and training effectiveness.
• Accurate HAR requires sophisticated analysis of human movement data.

Purpose of the Study:

• To investigate the impact of 3D data content on the classification accuracy of basic tennis strokes.
• To evaluate the effectiveness of combining player silhouette and tennis racket data for HAR.
• To determine the optimal input features for recognizing tennis strokes.

Main Methods:

• Utilized a motion capture system (Vicon Oxford, UK) to record 3D data.
• Employed the Plug-in Gait model with 39 markers for player body acquisition and a 7-marker model for the tennis racket.
• Applied an Attention Temporal Graph Convolutional Network for classification of tennis strokes (forehand, backhand, volley forehand, volley backhand).

Main Results:

• Achieved the highest classification accuracy of 93% when using both the player's entire silhouette and the tennis racket data.
• Demonstrated that incorporating racket position significantly improves the accuracy of human action recognition in tennis.
• Confirmed the importance of analyzing both whole-body movement and object interaction for dynamic actions.

Conclusions:

• The integration of 3D player silhouette and tennis racket data is highly effective for accurate human action recognition in tennis.
• Analyzing the complete player's body and racket position is essential for recognizing dynamic movements like tennis strokes.
• This approach offers a robust method for advanced sports analysis and performance evaluation.