Human Activity Recognition: A Comprehensive Survey of Machine Learning, Deep Learning and Attention Mechanisms

Human Action Recognition (HAR) has emerged as a significant area of research in artificial intelligence, driven by its wide-ranging applications, including video surveillance. The vast array of human activities in daily life adds complexity to accurately identifying these actions. Vision-based tasks such as action recognition and action prediction involve understanding and anticipating human movements from video data. Action recognition focuses on identifying human actions (current state) based on the full sequence of an activity. Recognizing human actions from video data has gained significant attention in recent years. A wide range of computer vision applications, including healthcare activity monitoring, detection of suspicious behavior, and human-computer interaction systems, rely heavily on this concept. The process of automatically identifying and labeling activities from sequences of video frames is referred to as activity recognition. In recent decades, considerable efforts have been undertaken to provide a strong and efficient system for recognizing and forecasting human actions. This paper provides a thorough review of recent developments in these fields. We systematically review existing models, widely used algorithms, technical challenges, prominent action recognition datasets, evaluation protocols, and discuss promising future research directions in detail.