Improving Human Action Recognition in Videos with Two-Stream and Self-Attention Module

Human action recognition has gained significant attention in recent years due to its potential applications in surveillance, human-computer interaction, healthcare, and security. Traditional approaches involve extracting low-level features from video sequences and using classifiers for labeling, while deep learning techniques have automated this process, providing more accurate and efficient action recognition. We propose a new approach for human action recognition in videos that employs a two-stream CNN to effectively capture appearance and motion information. Additionally, a self-attention network is used to selectively focus on the most informative regions of the video frames. Our method achieves an accuracy of 94.08% on the UCF101 dataset and demonstrates improved performance compared to existing techniques