Unsupervised Domain Adaptation for mmWave-based HAR via Text-mmWave Alignment

Human Activity Recognition (HAR) using millimeter-wave (mmWave) radar has emerged as a promising privacy-preserving and device-free sensing technology. However, its performance often degrades severely due to domain shifts when models trained in specific user, positional, or environmental settings are deployed in new scenarios. To address this challenge, we propose mmUDA, a novel unsupervised domain adaptation approach that employs textual semantics as a domain-invariant bridge for cross-domain alignment. In the source domain, mmUDA aligns radar-derived features with textual descriptions of activities, providing semantic grounding and enhancing feature generality. During adaptation, a teacher-student learning framework with pseudo-label generation and propagation progressively transfers knowledge from limited labeled source data to abundant unlabeled target data. To ensure reliability, we introduce an energy-based pseudo-label selection mechanism that filters out uncertain samples, mitigating noise accumulation in self-training. We implement and evaluate mmUDA on a commercial TI IWR1443BOOST radar platform across multiple users, positions, and environments. Experimental results demonstrate that mmUDA significantly improves cross-domain recognition accuracy, achieves robust generalization to unseen users and environments, and outperforms existing domain adaptation baselines. The code and dataset will be open-sourced after publication to facilitate further research in related fields.