A survey on deep learning architectures in human activities recognition application in sports science, healthcare, and security

In a typical human activity recognition (HAR) system, human activities are recognized by collecting data from inertial sensors (i.e., Inertial measurement unit (IMU)) or visual sensors (i.e., cameras). Then, the collected data is labelled with human activities. In turn, the data is used to train machine learning (ML) or deep learning (DL) models. HAR systems are widely used in different applications such as security, healthcare, the Internet of Things (IoT), and sports domains. The highest accuracy rates are achieved by DL models. In this context, we review the recent advancements of HAR systems in three trendy domains, namely, 1) sports science, 2) healthcare, and 3) security. The aim of this review is to reveal the most widely used DL architectures alongside the highest achieved accuracy rates in each of these domains. Both the Convolution Neural Network (CNN) and the Long Short Term Memory (LSTM) architectures achieved the best performance in both fields of sports science and healthcare. In the security field, the best performance was achieved by the adapted VGG-16 model.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022.