<p>Video data analysis identifies which, when, where, and by whom an activity is performed and extracts valuable information, insights, and knowledge. Videos, in contrast to photos, are inefficiently handled by a fixed-sized design architecture because their temporal duration varies greatly. Over the last three decades (1994–2024), various approaches based on feature representation, like traditional techniques and deep machine learning functionalities, have been developed to construct a reliable and precise framework and architecture for action recognition, particularly human action recognition for computer vision applications. Deep learning architectures have been developed to analyze and comprehend actions in video data. The field of human action recognition has shown tremendous development with the use of deep learning architectures, with a significant emphasis on enhancing accuracy, efficiency, and real-time capabilities. This systematic review comprehensively analyzes the evolution of deep learning architectures for human action recognition, evaluating their relative strengths across multiple dimensions including accuracy, computational efficiency, temporal modeling capability, and generalization. The review identifies key challenges when working with these architectures for video analysis and human action recognition, while providing evidence-based recommendations for selecting appropriate models based on specific application requirements. This paper will guide researchers in navigating the architectural landscape for human action detection and recognition from videos, offering insights into which approaches are best suited for particular contexts.</p>