Vision-Inspired Image Quality Assessment for Radar-Based Human Activity Representations

A research team led by Huy Trinh has published a paper introducing a novel framework for improving radar-based human activity recognition (HAR), addressing a critical gap in privacy-preserving monitoring technology. Radar-based systems using micro-Doppler spectrograms offer significant privacy advantages over cameras in sensitive environments like long-term care facilities, but their effectiveness has been limited by noise and clutter in the signals. The researchers benchmarked three existing denoising techniques—adaptive preprocessing, adaptive thresholding, and entropy-based denoising—while highlighting the shortcomings of conventional metrics in low signal-to-noise ratio conditions. Their work demonstrates how perceptual image quality measures can provide better assessment than standard error-based metrics for these radar representations.

The team's key innovation is a two-part framework that expands HAR beyond dynamic movements to include static activities. They developed a temporal tracking algorithm to enforce consistency across observations and a no-reference quality scoring algorithm to assess the fidelity of range-angle (RA) feature maps. Experimental results show these techniques significantly enhance both classification performance and system interpretability. This advancement opens the door for more reliable radar-based monitoring systems that can distinguish between activities like sitting, standing, and walking without compromising individual privacy through visual surveillance.