Wearable sensing and deep learning for human activity and action recognition in sport and physical activity: a systematic review

Abstract

The fast merging of sensors which can be worn, the interconnected network of things, and the power of artificial intelligence changed the game in measuring movements during sports and physical activity, however, the information on how such systems can detect various athletic activities is still scattered mainly within engineering, sports-science, and health fields. In this systematic review the authors gathered studies which used wearable or vision-based sensors with different forms of machine/deep-learning models to identify human actions and activities during sporting and physical activity scenarios. Scopus, PubMed, and ScienceDirect were searched in compliance with PRISMA 2020 and 658 records were found. After removing 10 duplicates, 648 were screened, 50 full texts were evaluated, and 20 studies were found eligible. These were peer-reviewed English-language publications or reviews published within the last five years (2021 to 2026). Main topics of the research are the combination of inertial and multimodal sensors with the attention-based and hybrid convolutional-recurrent architectures that have yielded a high level of recognition performance for sports actions, together with the trend in research focusing on free-living, edge-deployable, and personal monitoring. The review points to key areas for further research being generalisability, standardised reporting, and athlete-centred validation.