Author(s): SHEU, Y., CHEN, T.Y., JUANG, W.H., WU, S.K., Institution: NATIONAL FORMOSA UNIVERSITY, Country: TAIWAN, Abstract-ID: 518

INTRODUCTION:
Traditional table tennis training relies heavily on coaches' subjective experience, making it difficult to provide precise quantitative data. With the advancement of wearable sensors and AI, Edge AI offers new possibilities for real-time motion analysis. This study develops a TinyML-integrated intelligent racket capable of low-latency stroke recognition and VR visualization feedback without cloud dependency.
METHODS:
The hardware core utilizes an ARM Cortex-M7-based STM32F767 microcontroller paired with a high-performance ICM-20948 9-axis inertial measurement unit (IMU), precisely embedded within the racket handle. The system performs acquisition of acceleration and angular velocity signals for six core stroke techniques, including forehand/backhand drives, chops, and pushes. During the data processing phase, the raw sensor data were randomly partitioned into an 80% training set and a 20% testing set to evaluate the performance of four distinct deep learning architectures. Ultimately, an optimized and quantized Convolutional Neural Network (CNN) model was selected for deployment on the MCU. To ensure seamless feedback, classification results are transmitted instantaneously to a VR environment via an nRF24L01P wireless module.
RESULTS:
The study recruited four professional athletes and four amateur players for empirical testing. Initial statistics revealed that in a test of 20 strokes per person, the raw recognition rates were 83.75% for professionals and 85.58% for amateurs. To establish a more robust baseline, the research filtered high-standard stroke trajectories to perform deep model optimization. The refined CNN model demonstrated exceptional performance; the recognition accuracy for each individual stroke type exceeded 96%, while the overall average accuracy reached 98.2%. These results prove that the embedded AI racket possesses a rigorous and accurate classification standard. For general players, the system can effectively identify postural deviations and provide guidance through the VR interface, facilitating precise movement correction and autonomous skill acquisition.
CONCLUSION:
Compared to vision-based systems that rely on cameras, this sensor-based approach offers significant advantages, including immunity to ambient lighting conditions, enhanced privacy, and significantly lower computational overhead. Achieving such high accuracy confirms the robustness and feasibility of implementing TinyML on resource-constrained embedded devices. While the current results are promising, future research will focus on enhancing the model's generalization capabilities to accommodate users of different body types and strength levels, while expanding the database to include complex, multi-stage maneuvers. In conclusion, this study successfully integrates AI recognition with VR feedback, providing a high-efficiency, portable, and cost-effective cutting-edge solution for the future of personalized and digitally-augmented athletic education.