A lightweight hybrid vision transformer network for radar-based human activity recognition.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• Radar-based human activity recognition (HAR) provides privacy-preserving and lighting-robust sensing.
• Deep neural networks excel at classifying radar micro-Doppler signals for HAR but are computationally intensive for embedded systems.

Purpose of the Study:

• To develop an efficient and lightweight network for radar-based HAR that balances accuracy and computational cost.
• To address the challenges of implementing complex deep learning models in resource-constrained embedded applications.

Main Methods:

• Proposed a lightweight hybrid Vision Transformer (LH-ViT) integrating efficient convolutions with Vision Transformer (ViT) self-attention.
• Employed Feature Pyramid architecture for multi-scale micro-Doppler map feature extraction.
• Utilized stacked Radar-ViT with fold/unfold operations and RES-SE blocks to reduce computational load and enhance features.

Main Results:

• The proposed LH-ViT method demonstrated superior expressiveness and computing efficiency compared to traditional methods.
• Experimental results on two human activity datasets validated the network's performance advantages.

Conclusions:

• The LH-ViT offers an effective solution for radar-based HAR in embedded systems, achieving high accuracy with reduced computational requirements.
• This approach enables practical deployment of advanced HAR technologies in devices with limited resources.