Comparison of Machine Learning Algorithms for Ball Velocity Prediction in Baseball Pitcher using a Single Inertial Sensor

Abstract

Ball velocity of pitching is an important factor in baseball players. Commonly, ball velocity measurement requires specific devices such as radar gun. On the other hand, Gomaz et al. developed the accurate ball velocity measurement using two inertial sensors on pelvis and trunk. Recently, smartphone installed inertial sensor is popular device in daily life. Therefore, if ball velocity can be measured by only a single inertial sensor, baseball players can measure own ball velocity by only smartphone in daily life and various situations. Thus, the objective of this study is to propose and evaluate the ball velocity prediction method using the only a single inertial sensor. The proposed method predicts ball velocity using by a single inertial sensor and machine learning technique. Five machine learning algorithms (linear regression, support vector machine, gaussian process, artificial neural network, and M5P) predicted ball velocity by data of single inertial sensor, body height, and body weight. In this study,  Gomaz et al.’s public data for ball velocity and inertial data during pitching of baseball players were used for this investigation.  Sensor placement was either sternum or pelvis. Accuracy of prediction was evaluated by root mean square error (RMSE) between actual and predicted value via leave-one-out cross-validation. The results showed that greatest algorithm (M5P) could accurately predict ball velocity by only single inertial sensor and body parameters (RMSE < 2.0 mph). These results suggest that ball velocity can be measured by only single inertial sensor such as smartphone.