Whitening-Aided Learning from Radar Micro-Doppler Signatures for Human Activity Recognition

Area of Science:

• Artificial Intelligence
• Signal Processing
• Radar Technology

Background:

• Deep learning, particularly CNNs, is widely used for human activity recognition (HAR) with radar.
• Current CNN models for HAR often use batch normalization (BN) for training optimization and generalization.
• Radar micro-Doppler signatures are key inputs for these deep learning models.

Purpose of the Study:

• To introduce and evaluate whitening-aided CNN models for radar-based HAR.
• To demonstrate the benefits of replacing batch normalization layers with whitening layers.
• To improve the classification accuracy of human activities using radar sensors.

Main Methods:

• Replaced batch normalization (BN) layers in CNN models with whitening layers.
• Utilized whitening's ability to center, scale, and decorrelate activations.
• Exploited whitening matrices' rotational freedom to align latent space activations with activity classes.

Main Results:

• Whitening-aided CNN models achieved superior classification accuracy compared to BN-based models.
• Whitening effectively decorrelated radar micro-Doppler signature activations.
• The proposed method showed significant performance gains on real-world activity data.

Conclusions:

• Whitening-aided CNNs offer enhanced performance for radar-based human activity recognition.
• Whitening provides advantages over batch normalization by decorrelating features.
• This approach holds significant potential for improving HAR systems using radar sensors.