A Deep Learning Approach for Table Tennis Forehand Stroke Evaluation System Using an IMU Sensor.

Area of Science:

• Sports Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Home-based physical activity is crucial for mental well-being, especially during lockdowns.
• Table Tennis training can be enhanced with smart, privacy-preserving virtual coaching systems.
• Developing accurate performance evaluation for sports training is essential for skill improvement.

Purpose of the Study:

• To develop a performance evaluation system for Table Tennis Forehand strokes using a virtual coach.
• To compare the effectiveness of a proposed Long Short-Term Memory (LSTM) model against 2D Convolutional Neural Network (2DCNN) and Radial Basis Function Support Vector Regression (RBF-SVR) models.
• To analyze the impact of hyperparameters on machine learning model performance in evaluating sports movements.

Main Methods:

• Collected sensory data on racket movement and orientation from 16 players performing Forehand strokes.
• Developed and modified LSTM models for time-series analysis and performance evaluation.
• Compared the LSTM model's accuracy against 2DCNN and RBF-SVR using Root Mean Square Error (RMSE).

Main Results:

• The modified LSTM models demonstrated lower estimation errors compared to 2DCNN (33.79%) and RBF-SVR (4.24%).
• All nonlinear regression models showed good fit with the observed data.
• The LSTM model proved to be the most effective regression method for evaluating Forehand strokes.

Conclusions:

• The proposed LSTM model offers a powerful and accurate method for evaluating Table Tennis Forehand performance.
• Virtual coaching systems can effectively support home-based sports training and promote health.
• This technology has the potential to enhance skill development and maintain physical activity levels remotely.