mmWave Radar Activity Recognition Based on Dual-View Consistency Constraint

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China

Abstract：To address unstable predictions and insufficient cross-subject robustness of mmWave radar human activity recognition models under slight input perturbations, this paper proposes a robustness enhancement method based on dual-view consistency constraint. Physically plausible perturbations, including amplitude scaling, global phase shift, linear phase drift, temporal frame dropout, range masking, chirp masking, and additive Gaussian noise, are directly imposed on complex-valued radar samples to construct two stochastic views for each sample. A shared recognition backbone is then optimized with both supervised classification loss and Jensen--Shannon divergence based consistency loss, together with a dynamic weighting schedule. Experiments are conducted from the aspects of basic recognition performance, in-set/out-of-set generalization gap, multi-SNR anti-noise capability, and prediction stability under repeated perturbations. Results show that the proposed method consistently improves strict consistency and mean vote agreement on RT2DCNN, DT2DCNN, and RDT3DCNN pipelines, and also yields better out-of-set performance under selected medium-SNR conditions. These results demonstrate that the proposed method can effectively suppress prediction fluctuation and improve the robustness of mmWave radar activity recognition in perturbed scenarios.

Keywords： computer application mmWave radar activity recognition perturbation consistency