A Comparative Study of Machine Learning Algorithms for Padel Tennis Shot Classification

Abstract

En esta comunicación se presenta una comparativa de modelos de machine learning para clasificar distintos golpes de pádel. Para ello se creó la primera base de datos con golpes de pádel y se desarrolló un dispositivo para la detección de los mismos.

During the last few years, the use of MEMS in sport activity has become especially relevant. Among other different applications, data collected through MEMS have allowed the development of shot classifiers in some racket sports, such as tennis or table tennis. However, despite the astonishing worldwide growth of padel tennis, there are no previous attempts to generate a shot classifier for this sport. This woerk aims a) to create the first padel shot dataset in the literature, and b) to develop a shot classifier eith the capacity to distinguish 13 different padel shots. For the first purpose of the present study, a LSM9DSI IMU included in the Sense Hat of a Raspberry Pi collected data from 2328 strokes performed by 12 real players. Wearing the data acquisition devide in the dominant hand, each participant performed the shots in 13 different separate tests, each of them consisting of iteratively repeating a single type of shot. A developed algorith based on speed and acceleration signals provided by the IMU was used to identify the strokes and segment the data. Once the strokes had been singled out, five different Machine Learning algorithms, namely Neural Network (NN)m one-dimensional Convolutional Neural Network (1D CNN), decision tree, SVM, and K-Nearest Neighbors (K-NN), have been compared. The findings show that the 1D CNN performed better than the other algorithms achieving an accuracy of 93%.