IMU sensors enable human activity recognition (HAR) with continuous, privacy‑friendly monitoring. However, building models that generalize across diverse users and real‑world conditions requires large amounts of labeled IMU data, which are costly to collect. Existing approaches mainly rely on augmentation or synthesis, but indiscriminately adding virtual samples offers limited coverage and may introduce noisy supervision.
We therefore propose a coverage‑aware virtual IMU augmentation framework. First, in a learned sensor embedding space we select diversity anchors and scarcity anchors, convert each anchor's dynamics into prompts, and generate virtual IMU candidates accordingly. Next, we rank candidates by a selection cost that combines anchor proximity and label consistency. Finally, the selected virtual samples are incorporated into HAR training with reliability‑based weights.
Experiments on several public HAR benchmarks show that our method consistently improves recognition accuracy over strong baselines, and ablation studies confirm the effectiveness of anchor selection, the cost function, and the weighting scheme.
Review