Learning Three Dimensional Tennis Shots Using Graph Convolutional Networks

Area of Science:

• Sports Science
• Biomechanics
• Artificial Intelligence

Background:

• Human movement analysis is crucial in sports for performance assessment and training.
• Accurate recognition of fundamental tennis strokes like forehands and backhands is vital for quantitative game analysis.

Purpose of the Study:

• To evaluate the effectiveness of Spatial-Temporal Graph Neural Networks (ST-GCN) for recognizing forehand and backhand tennis shots.
• To compare the performance of ST-GCN using fuzzy versus non-fuzzy input data for movement recognition.

Main Methods:

• Utilized 3D motion capture data from tennis players performing forehand and backhand shots, including player and racket movements.
• Implemented Spatial-Temporal Graph Neural Networks (ST-GCN) for movement recognition.
• Compared two data input methods: with and without fuzzy logic applied to the input graphs.

Main Results:

• The study found that using fuzzy input graphs significantly enhanced the performance of ST-GCN.
• Fuzzy input graphs proved to be a superior method for recognizing forehand and backhand tennis shots compared to non-fuzzy inputs.
• The ST-GCN model demonstrated effectiveness in distinguishing between forehand and backhand movements.

Conclusions:

• Fuzzy data representation improves the accuracy of ST-GCN for tennis stroke recognition.
• Spatial-Temporal Graph Neural Networks offer a promising approach for advanced biomechanical analysis in sports.
• This research contributes to the development of AI-driven tools for sports performance analysis and training feedback.