Federated Continual Learning Framework for Robust mmWave Human Activity Recognition Under LOS–NLOS Domain Shifts

Millimeter-wave (mmWave) detection is a widespread human activity recognition (HAR) method. However, due to the mmWave characteristics, it is challenging to per-form HAR in non-line-of-sight (NLOS) environments. Before the prevalence of artificial intelligence (AI) technologies, mmWave-based HAR systems mainly relied on tradi-tional signal processing and handcrafted feature extraction (e.g. time–frequency analy-sis, multi-path modeling, micro-Doppler signature analysis), combined with classical classifiers such as Support Vector Machine (SVM) or Random Forest. However, these approaches were highly sensitive to environmental variations and failed to generalize in NLOS conditions. With the advent of AI, techniques such as Convolutional Neural Network (CNN) have played a crucial role in feature extraction from two-dimensional images generated by mmWave radar signals. In this paper, for the first time, a frame-work for Federated Continual Learning (FCL) based on CNN is proposed for improving radar detection accuracy in NLOS environments while preserving personal privacy through local training without uploading radar tensors and personal data to the global server for model aggregation. Additionally, the FCL model learns from both LOS and NLOS environments for improving cross-domain adaptability and recognition.