Parametric Spectral Attention for Real-Time mmWave Radar Pose Recognition from Raw ADC Data

Conventional mmWave perception pipelines decouple signal conditioning (e.g., fixed STFT) from feature learning. This disjoint approach often discards phase information and incurs high latency, limiting deployment on edge devices. This letter presents PSA-Mamba, a real-time framework operating directly on raw radar ADC data. We introduce a Parametric Spectral Attention frontend to bridge the gap between physical sensing and digital interpretation. Unlike deterministic windowing, this module synthesizes adaptive matched filters via backpropagation, maximizing the Signal-to-Clutter Ratio (SCR) for dynamic targets. Furthermore, a Bidirectional Mambaencoder is integrated to model long-range kinematic dependencies with linear computational complexity. Experimental results demonstrate a state-of-the-art accuracy of 96.5%. In addition, the proposed architecture reduces inference latency by 8× compared to Transformer baselines, offering a robust solution for resource-constrained embedded sensing.