Benchmarking ML and DL Techniques for Human Activity Recognition: A Comparative Study

Human activity Recognition (HAR) is a rapidly developing field with its applications in the field of healthcare, smart homes, fitness tracking and surveillance. HAR plays an important role in improving the functionality and intelligence of modern applications by enabling systems to identify and interpret human activities. Many strategies have been proposed over the time, ranging from conventional machine learning techniques to advanced deep learning methods and emerging framework designed for resource constrained environments. Recently, TinyML has created more opportunities for HAR implementation in resource constrained environment like wearable devices and Internet of Things (IoT) platforms where low battery consumption and computational efficiency are required. This study offers a thorough comparison of previous HAR techniques, assessing their effectiveness based on some parameters such as accuracy, complexity, hardware requirements and power consumption. After comparison, result shows that traditional ML models gradient boosting achieved accuracy of 86%, while DL model achieved higher accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 6 \%}$</tex>. DL models provided superior performance in recognizing human activities. In this paper we will also try to highlight the advantages and disadvantages trade off associated with each approach. Our findings highlight the increasing demand of TinyML techniques for edges-based application and also pointing out the key challenges and gap associated with this field. This Paper provides valuable insights for researchers aiming to design effective and scalable HAR systems for real world.