A machine learning driven approach for multivariate timeseries classification of box punches using smartwatch accelerometer sensordata

The objective of this study was the automatic classification of boxing punches to help athletes improving their punch performance. Our research proposes a system consisting of a smartwatch as the measuring device and serversided machine learning classification models. An app running on the wearable device measures the acceleration of the punches and sends the raw data to the backend where the classifcation is done. The classification result is announced to the athlete via a speech feedback. The system can classify four different classes of punch moves: frontal punch, hook, upper-cut, no-action. Additionally we classify which hand was used to punch and the athlete that performed the punch. In order to train our machine learning models, we created a dataset of more than 7600 punches performed by 8 different athletes. Using standard preprocessing and feature extraction methods the created algorithms reached an accuracy of 98.68% for the punch type classification and an accuracy of 90.80% for the hand classification and even an 80.54% accuracy classifing the athletes name. Via a developed dashboard web application the athletes can see their trainings statistics and use the system to optimize their punch performance. The main contributions of this paper are:1)We provide the first public available box punch dataset recorded with a standard smartwatch (https://kaggle.com/smartpunchteam/boxpunch-dataset)2)We treat box punch classification as machine learning problem and investigate different approaches on our dataset3)We provide the whole open source project on GitHub (https://github.com/smartpunch)