Corruption Robustness Analysis of Radar Micro-Doppler Classification for Human Activity Recognition

Abstract

Radar-based human activity recognition (HAR) is a popular area of research. Despite claims of high accuracy on self-collected datasets, the robustness of these models under data variations has been overlooked. This article focuses on corruption robustness analysis of radar micro-Doppler spectrogram classification for radar HAR. First, a taxonomy is proposed to classify corruptions into temporal, Doppler, and intensity domains, accompanied by strategies to effectively manage their severity for a balanced evaluation. Second, an analysis framework is presented to assess the robustness of corruption in radar sensing, providing insight into what factors to consider and how to evaluate using a dedicated corruption fmetric. Finally, a benchmarking study evaluates different model architectures and training methods to improve corruption robustness in two radar-based HAR tasks. The results indicate that higher capacity convolutional neural networks (CNNs) show improved classification accuracy, albeit with a risk of overfitting. In particular, adversarial training and data augmentation are identified as effective techniques to improve corruption robustness. However, corruption robustness is not a solved problem for the radar HAR task. Robustness to different types of corruption robustness could be dataset and model-dependent. In essence, our study contributes to a deeper understanding of the complex interplay between model architecture, training methods, and corruption robustness.