A Review of Human Action Recognition Based On Deep Learning

Human action recognition is a cornerstone of computer vision, with applications spanning from smart surveillance systems to assistive technologies. This survey paper provides a structured review of the two predominant approaches in the field: RGB-based and skeleton-based methods. The paper emphasizes the importance of these research avenues for advancing automated video analysis and understanding human behavior. We begin by reviewing RGB-based methods, which rely on color and texture information extracted from video frames. The survey discusses the evolution from handcrafted feature extraction techniques to the current deep learning paradigm. Convolutional neural networks (CNNs) have been pivotal in this transition, enabling the automatic learning of discriminative features from raw pixel data. Subsequently, the survey turns to skeleton-based methods, which focus on the temporal sequences of human joint locations. These methods offer robustness against view-point changes and clothing variations. We explore the use of recurrent neural networks (RNNs), particularly long short-term memory networks (LSTMs), for modeling the sequential nature of skeleton data. Additionally, we examine graph convolutional networks (GCNs) applied to the human skeleton as a graph structure. In conclusion, the survey reflects on the progress and challenges within each approach.