Adaptive Kinematic Masking for Urban Human Activity Recognition Using UWB Radar: A Sim2Real Approach

Ultra-wideband (UWB) radar is promising for human activity recognition (HAR), but its coarse micro-Doppler (m-D) resolution and the scarcity of labeled real-world data hinder cross-environment generalization. To address these challenges, we present a task-specific simulation-to-reality (Sim2Real) framework that integrates complementary micro-Doppler representations, adaptive kinematic supervision, and parameter-efficient domain adaptation. The generated masks supervise a lightweight kinematic mask (LKM) decoder during both simulated pre-training and few-shot real-world fine-tuning, encouraging motion-aware representation learning while allowing the auxiliary decoder to be removed at inference. Experiments on measured UWB radar data demonstrate improved cross-environment generalization without additional inference overhead from the auxiliary branch.