Human action recognition method based on Motion Excitation and Temporal Aggregation module.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition is crucial for intelligent systems.
• Existing methods struggle with low modeling efficiency and temporal feature loss.
• Effective temporal modeling requires capturing multi-state and multi-scale motion dynamics.

Purpose of the Study:

• To propose an efficient and effective human action recognition method.
• To address the limitations of low modeling efficiency and feature loss in temporal modeling.
• To enhance the capture of multi-state and multi-scale temporal information.

Main Methods:

• Introduced the Motion Excitation and Temporal Aggregation (META) module.
• META comprises Multi-scale Motion Excitation (MME) and Squeeze and Excitation Temporal Aggregation (SETA) modules.
• Employed temporal relational sampling, cross-modality pre-training for optical flow, and combined spatiotemporal features.

Main Results:

• Achieved 96.0% accuracy on the UCF101 dataset.
• Achieved 71.2% accuracy on the HMDB-51 dataset.
• Demonstrated superior performance compared to contemporary methods.

Conclusions:

• The proposed META module significantly improves human action recognition.
• The method effectively captures multi-state and multi-scale temporal information.
• META offers a promising approach for efficient and accurate action recognition.