Radar Signal Processing and Its Impact on Deep Learning-Driven Human Activity Recognition

Area of Science:

• Radar Signal Processing
• Machine Learning
• Computer Vision

Background:

• Human activity recognition (HAR) is crucial for smart security and healthcare.
• Radar technology offers a privacy-preserving alternative for HAR.
• Integrating advanced signal processing with deep learning is key to improving HAR.

Purpose of the Study:

• To investigate the integration of CNNs with radar signal processing for enhanced HAR.
• To evaluate different radar map generation techniques and CNN architectures.
• To identify optimal configurations for real-time, resource-constrained HAR applications.

Main Methods:

• Utilized three 2D radar processing techniques: range-FFT time-range maps, STFT time-Doppler maps, and SPWVD maps.
• Evaluated four CNN architectures: VGG-16, VGG-19, ResNet-50, and MobileNetV2.
• Analyzed twelve CNN and preprocessing configurations for accuracy and computational efficiency.

Main Results:

• MobileNetV2 with STFT preprocessing achieved 96.30% accuracy.
• This configuration demonstrated high computational efficiency with a 2.57 ms inference time.
• Spectrogram generation time was 220 ms, suitable for real-time processing.

Conclusions:

• Radar-generated maps serve as effective visual data for HAR, ensuring privacy.
• The study highlights the trade-offs between preprocessing complexity and recognition accuracy.
• Optimal configurations, like MobileNetV2 with STFT, enable wider application of radar-based HAR in edge computing and AR.