A Comprehensive Review of Recent Deep Learning Techniques for Human Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition is a critical area in computer vision.
• Deep learning methods have significantly advanced action recognition capabilities.
• Recent advancements include the exploration of transformer architectures.

Purpose of the Study:

• To provide a comprehensive survey of deep learning-based human action recognition methods using RGB video data.
• To categorize recent methods into five distinct groups for clarity.
• To review emerging convolution-free, transformer-based approaches.

Main Methods:

• Categorization of deep learning methods into 2D Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), 3D CNNs, and multi-stream approaches.
• Inclusion and analysis of recent pure-transformer architectures (convolution-free).
• Performance comparison on four popular benchmark datasets and review of 26 benchmark datasets.

Main Results:

• Transformer-based methods demonstrate state-of-the-art results, outperforming traditional convolutional networks in many cases.
• Comparative analysis highlights the strengths and weaknesses of different deep learning architectures for action recognition.
• The survey provides a structured overview of 26 benchmark datasets.

Conclusions:

• Deep learning, particularly transformer architectures, offers significant advancements in human action recognition.
• The survey consolidates current knowledge and categorizes diverse methodologies.
• Future research directions are identified to guide further progress in the field.