Author(s): SHEU, Y.H., CHEN, T., WU, S., Institution: NATIONAL FORMOSA UNIVERSITY, Country: TAIWAN, Abstract-ID: 522

INTRODUCTION:
With the rise of sports technology and Artificial Intelligence (AI), the integration of sensors with AI is gaining popularity in the sports industry. The emerging technology of Tiny Machine Learning (TinyML), developed in recent years to address the limited computational resources of devices, is being widely adopted. This study introduces a cost-effective The Real-time Table Tennis Swing Classification System Based on TinyML. The system aims to reduce the training burden on table tennis coaches or players and enhance the interest of beginners in learning professional techniques.
METHODS:
The system is primarily comprised of two main components: the hardware, known as the embedded AI smart table tennis racket, and the software, referred to as the computer application. The hardware aspect, the embedded AI smart table tennis racket, includes the racket itself and an RF wireless receiver for receiving measurement signals. The software component, the Computer Application, is responsible for displaying the classification results. This combination establishes a real-time table tennis swing posture classification system based on TinyML technology.
After collecting the swing data, it is transmitted to the AI-MCU (Microcontroller Unit) on the Embedded AI Smart Table Tennis Racket. The AI-MCU uses its TinyML model for real-time swing posture classification. The classified swing data is then wirelessly transmitted via Radio-Frequency (RF) to the computer-side application for display. The embedded AI smart table tennis racket utilizes the STM32F7 component as the main control chip. The STM32F7 component is responsible for communication and control of 6-axis sensorts and performs processing and calculations for the TinyML model.
RESULTS:
In the data analysis, real-time classification is achieved using AI deep learning theory. By collecting swing data from our schools players, four different AI deep learning models are tested for data segmentation and classification accuracy of swing data. The experimental results show that Convolutional Neural Networks (CNN) have the best capability for waveform segmentation and swing classification accuracy. In practical swing tests, the converged training data and optimized CNN model achieve an average classification accuracy of approximately 98.3% for actual table tennis player swings. Through the development of the Embedded AI Smart Table Tennis Racket, it not only enables swing posture classification but also corrects the swing postures of typical table tennis beginners, achieving the intelligent training goal for table tennis.

CONCLUSION:
Through the development of the Real-time Table Tennis Swing Classification System Based on TinyML, it not only enables swing posture classification but also corrects the swing postures of typical table tennis beginners, achieving the intelligent training goal for table tennis. In the future, this research can also be combined with AR/VR related applications.