Radar-based human activity recognition with adaptive thresholding towards resource constrained platforms.

Area of Science:

• Engineering
• Computer Science
• Biomedical Engineering

Background:

• Radar systems offer privacy-preserving, contactless human activity recognition, unaffected by lighting.
• Current complex algorithms hinder deployment on resource-constrained embedded platforms.

Purpose of the Study:

• To develop an efficient method for highlighting regions of interest in multi-domain micro-Doppler signatures.
• To reduce computational cost and enable deployment of radar-based activity recognition on embedded systems.

Main Methods:

• An adaptive magnitude thresholding approach was developed to identify salient regions in micro-Doppler signatures.
• The method focuses on simplifying calculations for reduced computational load.

Main Results:

• The adaptive thresholding achieved up to 93.1% accuracy for recognizing six human activities.
• Outperformed deep learning methods by over tenfold in training time and memory footprint.
• Achieved twofold reduction in inference time compared to deep learning implementations.

Conclusions:

• The proposed adaptive thresholding method offers a computationally efficient solution for radar-based human activity recognition.
• This approach facilitates the deployment of advanced human activity recognition systems on embedded platforms.