Classification of pitch types outside the laboratory or game environment provide benefits in designing and outlining training routines to reduce injury and improve performance. Current classification approaches are based on vision-based optical data, however these optical data is often not available in training sessions. The aim of the current study is to use machine learning algorithms to classify pitch types, i.e., fastball, change-up and curveball, based on pelvis and trunk IMU data in training sessions. A total of 406 successful pitches thrown by 19 pitchers were used to classify pitch types. Three conditions were tested and evaluated in a binary (fastball versus others) and multi-class (fastball versus change-up versus curveball) classification approach. Condition A included the direct output of the PitchPerfect software, whereas condition B included features normalized by fastball characteristics. Condition C is a combination of both condition A and B. The random forest algorithm demonstrated the best predictions in both the binary and multi-class classification approach based on the highest accuracy and F1 scores, i.e., the harmonic mean between sensitivity and precision. Therefore, the random forest algorithm is the best fit for classifying pitch types based on pelvis and trunk IMU data in training sessions. In addition, the performance of the classification algorithm improved when using a binary classification approach. There were no relevant improvements when using additional features. The random forest algorithm can directly be implemented in the PitchPerfect application. Pitchers can use the pitch type data to track and tune their performances, whereas coaches can use the data to design match and training routines. Future research should focus on larger datasets, i.e., more pitchers and pitches, to pre-classify pitchers with similar pitching characteristics to improve the classification algorithms.

The purpose of this study was to determine whether a change of an accuracy restriction (AR) leads to a change in the separation time between the peak rotational velocities of the pelvis and the trunk during a pitching motion, while the ball velocity is kept constant. The peak rotational velocities of the pelvis, trunk, the separation time and ball velocity were measured on ten players (age: 20.4±2.4 years, length: 187.6±4.7 cm, bodyweight: 81.3±9.7 kg, pitching experience: 10.7±4.2 years). They performed two tasks twice. First, they had to perform ten fastballs on a large target with a low accuracy restriction. Secondly, they had to perform ten fastballs on a small target with a high accuracy restriction. A week later they performed the same tasks but in reversed order. The associations between the kinematic variables and the accuracy restriction were separately determined using generalized estimating equations. The rotational velocity of the pelvis was significantly associated with the accuracy restriction. The separation time and rotational velocity of the trunk were not significantly associated with the accuracy restriction. It cannot be concluded that the separation time was affected by a change in accuracy restriction.

One interesting part of the application of human activity recognition is sports motion recognition and classification. In recent years, many commercial wearable devices have been used for recording and supervising motion data information during sports. However, their claimed high-accuracy results but motion recognition and classification method have not been proven. This thesis project presents work special related to tennis stroke detection and classification. An automated and comprehensive tennis stroke recognition and classification method based on the inertial measuring unit sensor (accelerometer and gyroscope) and machine learning algorithm (Support vector machines) was proposed in this study. Seven tennis players with a different level of tennis skills were tested and recorded using a self-made IMU sensor system with four sensors (forearm, upper arm, trunk, and pelvis). Video footage from Playsight was manually notated as the golden standard for stroke type identification. SVMs was constructed to train the classification model to classify true shots to eight types of tennis strokes from the IMU signals. Across leave-one-out seven-fold cross-validation, the SVMs classification models were trained with data from a single IMU sensor on the forearm and upper arm with the prediction accuracies of 0.69 and 0.70 respectively. And further, both SVMs models were trained by enlarged training data, resulting in improved prediction accuracies of 0.75 and 0.77. Noticeably, the best prediction accuracy was achieved by training the SVMs classification model with fused data from the previous two sensors and with the enlarged training data. The final prediction result was 0.79. Even though there exist deficiencies such as skill level different of subjects, insufficient training datasets which may lead the results of validation and prediction less credible, the IMU sensor and SVMs machine learning algorithm still played well in the tennis stroke classification task. And we expect to have better accuracy results by feeding enough training data and using data-fusion combination of different IMU sensors to the upper extremity to SVMs classification model in future work.