A histogram transformer approach using attention-based 3D residual network for human action recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video action recognition is crucial for AI-driven applications.
• Existing methods often face challenges with computational complexity and accuracy.
• Lightweight and efficient models are needed for real-time video understanding.

Purpose of the Study:

• To propose a lightweight yet accurate video action recognition framework.
• To improve spatiotemporal modeling granularity and feature selection.
• To reduce computational redundancy in 3D CNNs.

Main Methods:

• Integration of 3D Convolutional Neural Networks (CNNs) with Histogram Transformer Blocks (HTB) and Split-Attention Residual Blocks (SAB).
• Application of Spatiotemporal Tensor Factorization (ST-Factor) to decouple 4D convolution kernels.
• Leveraging local statistical features via HTB and dynamic channel re-weighting via SAB.

Main Results:

• The proposed method achieves state-of-the-art (SOTA) recognition accuracy on UCF101/HMDB51 datasets.
• Maintained real-time inference speed, demonstrating efficiency.
• Significant reduction in computational redundancy through ST-Factor.

Conclusions:

• The developed framework offers a new paradigm for efficient and accurate video understanding.
• The integration of HTB, SAB, and ST-Factor provides a robust solution for lightweight video action recognition.
• This research contributes to advancing real-time video analysis capabilities.