Abstract

This review systematically examines recent advances in Human Activity Recognition (HAR) enabled by deep learning techniques, which play a critical role in improving human–machine interaction across applications such as healthcare, surveillance, and intelligent environments. The study analyzes peer-reviewed publications published between 2014 and 2024 to trace the evolution of HAR from traditional approaches to state-of-the-art deep learning models. It provides a comprehensive overview of commonly used datasets for both vision-based and sensor-based HAR systems, as well as deep learning architectures including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and hybrid models. The findings indicate a clear shift toward sensor-based HAR systems due to the flexibility and non-intrusive nature of wearable and ambient sensors. Moreover, the review highlights the capability of deep learning models to handle complex and unstructured data, resulting in notable improvements in recognition accuracy. Finally, this study identifies key research challenges and future directions, including data privacy concerns, model generalization, and the optimization of deep learning models for deployment on resource-constrained devices. Overall, this review offers a structured analysis of datasets, methodologies, and performance trends, providing valuable insights and a clear roadmap for future research in the HAR field.