End-to-End Hand Gesture Recognition on LowPerformance mmWave Radar via Complex-Domain Attention and 1D-Inception Transformer

Most mmWave radar gesture-recognition pipelines still follow a signal-processing-centric workflow-complex and typically followed by computer-vision or point-cloud postprocessing. Such pipelines are limited by the intrinsic resolution of the DFT and by hand-crafted, non-optimal feature transforms; moreover, on lower-end radars they often fail to produce clean spectrograms or point clouds.To address these issues, we introduce CuAttenFormer, an end-to-end, data-driven network that learns task-optimal transforms directly from mmW ave measurements. On a strong baseline, CuAttenFormer delivers an absolute gain of $\mathbf{1. 7 4 2}$ percentage points, demonstrating its superior effectiveness.