Abstract
Human Activity Recognition (HAR) is an active research topic as it finds use in many real-world applications such as health monitoring and biometric user identification. Smart wearables which form an integral part of the Internet of Medical Things (IoMT) and Cyber-Physical Systems can provide information about human activities on a daily basis, which may be used as soft biometrics for user identification. Over the last few years, one of the popular problem-solving approaches for HAR has been in the form of artificial intelligence methods. Since security is related to robustness, our primary aim is to solve the problem with better model capabilities. In this study, we consider machine learning algorithms like Random Forest (RF), Decision Trees (DT), K-Nearest Neighbors (k-NN)(and deep learning algorithms such as Convolutional Neural Networks (CNN), Long Short Term Memory (LSTM), and Gated Recurrent Units (GRU)) for the purpose of HAR. In order to improvise the model performance, we introduce optimization techniques along with CNN, LSTM, and GRU. We rely on Stochastic Gradient Descent (SGD), and optimizers Adam and RMSProp, and evaluate the strength of the models using Accuracy and F-1 score. Moreover, the study has been carried out on three datasets that incorporate several human activities. Our study indicates that adding a component of optimization increases the model performance, and the highest accuracy achieved in the study is almost 98%.