Distortion-resistant dual-branch millimeter-wave radar gait recognition for underground overhead-view scenarios

As a long-range, noninvasive behavioral biometric, gait is difficult to forge, requires no active cooperation, and is suitable for continuous monitoring. Millimeter-wave radar-based gait recognition provides a new technical approach for contactless personnel identification in complex underground environments. A high-mounted overhead-view millimeter-wave radar configuration offers greater safety and stability in practical deployments but introduces nonlinear projection in the overhead view, causing severe distortion and feature aliasing in micro-Doppler time-frequency spectrograms and substantially reducing gait recognition accuracy. To address this problem, a distortion-resistant dual-branch millimeter-wave radar gait recognition network, StripGait, was proposed for underground overhead-view scenarios. The network used a full-frequency vertical-strip partitioning strategy to strictly preserve the inherent physical coupling among micro-Doppler components; decoupled the global spatial envelope from local micromotion through a dual-branch architecture; and incorporated a Multi-Scale Temporal Fusion (MSTF) module to accurately decouple and map dynamic evolution patterns over different physical spans in micro-Doppler spectrograms through stepped temporal receptive fields. To address increased intra-class variation and blurred inter-class boundaries under overhead-view conditions, a hard-sample joint metric optimization strategy was introduced, and a composite metric constraint comprising ArcFace loss and hard-sample triplet loss was constructed, effectively reducing distortion-amplified intra-class dispersion and increasing inter-class decision margins. On a self-collected 45° overhead-view millimeter-wave radar gait dataset, StripGait achieved an average recognition accuracy of 94.5% with a single-frame inference time of 12.35 ms, striking a good balance between accuracy and computational efficiency and significantly outperforming existing mainstream spatiotemporal networks. It provides a highly robust technical approach for efficient personnel sensing in complex underground spaces.