Comprehensive Tennis Serve Training System Based on Local Attention-Based CNN Model

A high quality serving ability is crucial for winning a tennis match. Traditionally, tennis training has relied on the use of expensive and difficult to setup multiple highspeed cameras and computer vision technology. A more accessible, cost-effective, and user-friendly solution for tennis training is necessary. A comprehensive motion recognition system based on inertial sensors is proposed for tennis serve training in this article. The proposed system is designed to assist in tennis serve training by utilizing multiple inertial sensors to collect the player’s motion data. It employs small datasets data preprocessing and local attention based convolutional neural network (CNN) method for recognition. The system’s functionality includes landing location prediction, ball speed prediction, and the recognition of faulty actions in the serve action. We conducted comprehensive experiments to evaluate the performance of the proposed model. Compared with similar studies, the system in this article has high performance, with an average error of 0.72 units for the landing point prediction, an average error of 4.36 km/h for the ball speed prediction, and an accuracy rate of 98.2% for faulty actions recognition. Our work enables players to train efficiently and effectively, independent of environmental factors, and improve their movements using quantitative serve results and training recommendations.