Estimation of Upper Limb Kinematics in Baseball Pitching Using Sensor-Embedded Ball

We estimated the kinematics of the upper limb during baseball pitching using a sensor-embedded ball. Twenty collegiate pitchers performed 24 trials each, throwing fastballs, curveballs, and sliders. Kinematic data were collected using motion capture system and a sensor-embedded baseball. Machine learning models, including Linear Regression, Lasso, Random Forest, Gradient Boosting, and Support Vector Regression (SVR), were used to estimate joint angles. The evaluation metric was the R2 score from 10-fold cross-validation. Random Forest (R2 score (the average of all joint movement) = 0.85) and Gradient Boosting showed high accuracy (R2 score (the average of all joint movement) = 0.85), particularly for shoulder and elbow joints. The ensemble model further improved accuracy. The model demonstrated high accuracy in estimating joint angles for the shoulder joint (external/internal rotation) (R2 score (the average of all time) = 0.97) and elbow joint (supination/pronation) (R2 score (the average of all time) = 0.96). In the future, the application of this model is expected to facilitate the acquisition of kinematic data of the pitching arm in competitive environments.