Radar-Based Open-Set Human Activity Recognition With Distribution-Regularized One-vs-All Learning

Abstract

As Internet of Things (IoT) systems increasingly require privacy-aware sensing, radar-based human activity recognition (HAR) has become a promising option for smart homes, healthcare, and security. However, most existing methods assume a closed set of activities and often misclassify previously unseen behaviors. To address this issue, a distribution-regularized one-vs-all prototype learning (DROVA) framework is proposed for open-set radar-based HAR. First, frequency-guided decomposition enhances micro-motion cues and suppresses background interference, while deformable convolution adaptively aligns the receptive field with frequency-aware features. Second, a distribution-regularized one-vs-all objective jointly learns class prototypes and class-specific radii, thereby improving feature compactness and class separability. Third, during inference, a distance-based prototype decision rule accepts samples whose features lie within the learned class regions and rejects those outside. Experiments on real-world radar data show that DROVA achieves 89.33%  ±  0.69% KAcc, 89.59%  ±  1.44% UAcc, and 89.64%  ±  0.28% F-measure over five runs. Compared with SCRL, which yields the strongest baseline F-measure on the self-collected dataset, it improves UAcc by 6.88 percentage points and F-measure by 1.92 percentage points. On the Glasgow Human Activities dataset, DROVA also attains the best F-measure in both the six-subset and held-out NG Homes settings. These results suggest that the proposed framework is a promising and practically effective approach for open-set radar-based HAR.