Radar-Based Cross-Domain Human Behavior Recognition Using Physics-Informed Data Augmentation and Multi-Source Domain Adversarial Neural Network

Falling is a common but fatal human behavior in life. With the rapid growth of the aging population, fall-related human behavior recognition has been extensively investigated using radar. Nevertheless, human behavior recognition frequently exhibits suboptimal generalization capabilities due to the scarcity of labeled data. To address this problem, we propose physics-informed data augmentation (PIDA) for the micro-Doppler signature (m-DS) produced from frequency-modulated continuous wave (FMCW) radar to facilitate domain generalization (DG) of human behavior recognition. Specifically, leveraging classical electromagnetic wave propagation theory and radar signal processing principles, we introduce a PIDA strategy: 1) Distance Data Augmentation (DDA), which simulates electromagnetic attenuation effects across varying detection ranges, and 2) Behavior Pattern Data Augmentation (BPDA), designed to preserve intrinsic kinematic features while diversifying motion styles. Subsequently, combined with the proposed PIDA, a multi-source domain adversarial neural (MSDAN) network based on transfer learning is constructed to capture invariant features of human behaviors. To verify the superiority of PIDA, a series of experiments are conducted on various unseen domain data. The results showed that the PIDA method improves average DG accuracy by 1.52%, 2.48% and 4.88%, respectively, on the own dataset and 3.67%, 6.29% and 9.78%, respectively, on the public dataset, compared with baseline data, traditional optical data augmentation (TODA), and generative data augmentation (GDA) methods. Compared with current fall-related human behavior recognition works, our proposed method has a small DG gap.