Human Action Recognition Using CNN and ViT

Human Action Recognition (HAR) plays a significant role in computer vision, which has an extensive applications across various domains, including human-machine interaction, healthcare monitoring, and entertainment. This paper is aimed at investigating and comparing the performance of Convolutional Neural Networks (CNN) and Vision Transformers (ViT) for HAR from image datasets. CNN are prominent for their ability to capture localized spatial features through hierarchical layers of convolution, making them highly effective in tasks requiring detailed, fine-grained feature extraction. In contrast, self-awareness mechanisms are employed by ViT to model global relationships within images, thus providing a broader and more comprehensive understanding of the entire scene. Through a series of thorough experiments, this paper evaluates the respective strengths and limitations of both models. The results indicate that while CNN exhibit faster initial learning due to their proficiency in localized feature extraction, ViT, although slower in the early stages, eventually surpass CNN by excelling in global feature extraction and scene-level interpretation, leading to improved overall performance.