Object sensors are widely used for motion capture, particularly in sport motion analysis for classification of strokes. In this paper, a comparative study was performed to examine the Forehand strokes classification of three Machine Learning (ML) models in Table Tennis. Support Vector Machine (SVM) with Radial Basis Function (RBF) kernel function, Long Short-Term Memory (LSTM) and 2-Dimensional Convolutional Neural Network (2D-CNN) are considered. Tuning the models' parameters and examining sensitivities of the models to the number of training datasets and type of data are studied. All models were trained and tested on Table Tennis Forehand strokes' signals, which were collected from professional and novice Table Tennis players'. A BNO055 sensor was attached to the center of the standard racket as an Object sensor to measure the strokes' signals. To provide a robustness assessment of all models to type of dataset, the study employs 80% of novice players' strokes samples for testing classification accuracy. The empirical results suggest that the LSTM and the 2D-CNN classification outperforms with a substantial performance increase of approximately 7% more than the RBF-SVM. In addition, in the case of models sensitivity to the number of training samples it can be said that, the LSTM model performance on the self-collected dataset is not significantly sensitive to the number of training samples. However, the number of the developed RBF-SVM parameters is significantly less than other two deep models and are easy to set. Study shows that the modified LSTM model performed better than other two models.
© Copyright 2020 IEEE Sensors Journal. IEEE Institute of Electrical and Electronics Engineers. Alle Rechte vorbehalten. / Übersetzt mit https://www.DeepL.com/Translator