IoT smart athletics : Boxing glove sensors implementing machine learning for an integrated training solution

Abstract: It is very common in everyday life for people to use data generated by sensors like accelerometers and gyroscopes, whether they are on the mobile phone, smartwatch or other smart devices, for analysis of their movement or tracking their habits. This study is focused on boxing, and proposes a test where the generated data are put through machine learning algorithms in order to output information on the type of punch thrown by the user. Furthermore, the possibility of implementing ML on Android is examined. This thesis was performed by conducting a literature study, and an experimental study. For the literature study, researches similar to this were examined to gather information and insight on what the most common practices are, regarding the setup of the device used to collect the data, both in terms of sensor placement on the body and sensor setup like the optimal data output rates. The experimental part was conducted using custom hardware implementing an accelerometer and a gyroscope in which the wearer of this device would proceed to throw 6 types of punches (jab, cross, left & right uppercut, and left & right hook) to generate the data to be analyzed. It was technically possible to use Android for ML, but it was the least optimum way to execute the algorithms, so a PC was used instead. After putting the data through multiple ML algorithms, the results show that with our current hardware set up it was not possible for the ML algorithms to adequately classify the type of punches with mediocre accuracy scores ranging from 37.37% - 59.16% depending on the algorithm.