Beyond Radial Limits: Recovering Direction-Invariant Walking Dynamics for Robust mmWave Radar Person Re-Identification

Person re-identification (ReID) is essential for applications such as access control and smart environments. While vision-based ReID has achieved impressive performance, it is vulnerable to illumination, clothing changes, and privacy concerns. Millimeter-wave (mmWave) radar offers a privacy-preserving and robust alternative, but practical radar-based ReID remains challenging due to sparse and noisy measurements as well as strong view-dependent distortions caused by radial sensing. In this paper, we propose mmReID, a radar-native gait-based ReID system that revisits feature representation and modeling from the perspective of mmWave sensing. Instead of converting radar point clouds into Cartesian coordinates, mmReID operates directly in the polar domain, preserving the inherent sensing structure and avoiding noise amplification induced by angular uncertainty. To address walking direction-dependent gait distortion, mmReID incorporates a radial perspective alignment strategy that maps observations from arbitrary walking directions into a predefined radial reference direction using a learned affine transformation operator, yielding consistent gait features. We construct three datasets across three indoor scenes with 57 subjects under diverse real-world conditions. Extensive experiments using a commercial mmWave radar demonstrate that mmReID achieves state-of-the-art performance and strong generalization.