To ensure health benefits and prevent injuries, the correct execution of fitness exercises is essential, particularly when vulnerable individuals are involved, such as during rehabilitation. As it is difficult for a person to assess the execution quality for themselves and most people cannot afford a personal trainer at all times, an automated assessment of execution quality is desirable. Whereas human activity recognition with modern sensor technologies has become a fundamental topic in scientific research and industry over the past decade, the execution quality of exercises is rarely addressed. In this paper, we assess the applicability of machine learning-based classification to differentiate not just between different fitness exercises, but also their execution quality. For this purpose, we propose three different system variants to recognize three different fitness exercises and at least three typical execution errors each based on acceleration and gyroscope data from up to four body-worn sensors. In our evaluation, we utilize data we recorded from 16 different participants to determine our systems’ recognition performance for different application and implementation scenarios.

Preface: "We initially planned to host JCSG 2021 at Staffordshire University, Stoke-on-Trent, UK as a hybrid conference with delegates joining in person or online. However, due to the ongoing global situation with COVID-19, JCSG 2021 was organised as a virtual event [...]."