Radar-based human activity recognition (HAR) is a popular research field. Despite claims of high accuracy on self-collected datasets, the ability of these models to handle unexpected scenarios has been largely overlooked. This work introduces a framework for analyzing corruption robustness of radar micro-Doppler spectrogram classification. A set of corruptions are categorized, applied, and systematically tested on common model architectures. Diverse training methods, including adversarial training, cadence velocity diagram (CVD) transformation and data augmentation, are explored. The performance is evaluated on two tasks: indoor HAR and continuous aquatic HAR. Our study unveils several insights. Firstly, relying solely on accuracy may not adequately assess model performance due to dataset limitations. All well-trained models exhibit sensitivity to corruptions. Secondly, deeper convolutional neural network (CNN) models excel in both accuracy and robustness, but confront the problem of overfitting to background. Thirdly, adversarial training enhances robustness against corruptions, albeit at the cost of a slight decrease in accuracy. Lastly, combining data augmentation and adversarial training achieves a balance between accuracy and robustness. In essence, our study contributes to a more profound understanding of the complex interplay between model architecture, classification accuracy, and corruption robustness in radar HAR tasks.

Radar-based human activity recognition (HAR) is a popular area of research. In this paper, we investigate methods to improve the generalization of micro-Doppler-based swimming activity recognition. We identify three main challenges to this task: a small dataset lacking motion diversity, inaccurate period estimation, and inefficient network design that does not take into account the unique characteristics of spectrograms. To address the limited motion diversity, we propose a spectral data augmentation tailored for micro-Doppler spectrograms, including positive augmentations that account for physical fidelity and negative augmentations that penalize the unrealistic examples. We also investigate self-supervised pre-training to effectively use these negative augmentations. To address inaccurate period estimation, we introduce a segmentation approach based on energy distribution to handle temporal period variation. To exploit the spreading pattern of limb motion in the Doppler dimension and the continuous properties of torso motion in the temporal dimension, we design a module consisting of both 2D convolution and 1D temporal dynamic convolution to serve as a feature extractor. Our evaluation on a self-collected swimming activity recognition dataset shows that our model achieves the best classification accuracy and robustness to corruptions, even compared to much larger models and multi-domain fusion models.

Radar-based Human Activity Recognition (HAR) is popular because of its privacy and contactless sensing capabilities. However, a major challenge in this area is the lack of large and diverse datasets. In response, we present a novel framework that uses generative models to transform textual descriptions into motion data, thereby simulating radar signals. This approach significantly enriches the realism and diversity of the dataset, especially for infrequent but critical activities such as falls and abnormal walking. Textual descriptions capture the semantic complexity and ambiguity of actions, thereby improving intraclass diversity. Our framework scales the data generation process and improves simulation fidelity by controlling gait variation, multi-viewpoint adaptation and background noise modelling. The simulated micro-Doppler dataset can be used for model comparison and transfer learning to improve recognition in realworld data, even when available data samples are scarce. Our approach significantly mitigates the challenge of data shortages, enabling significant advances in activity recognition with limited samples.