The recognition of human activities using WiFi Channel State Information (CSI) facilitates contactless, long-range, and visual privacy-preserving sensing in confined indoor environments. However, the strong environmental dependence inherent to CSI presents a challenge for robust cross-domain generalization, limiting its practical applicability. Drastic environmental variations, such as transitions between line-of-sight (LOS) and non-line-of-sight (NLOS) through-wall scenarios or changes in antenna configurations, introduce a significant domain gap that can lead to severely degraded model performance at test time. To address the challenge of model generalization in these demanding cross-scenario and cross-system settings, an area that remains under-explored, this work investigates the effectiveness of data augmentation techniques commonly utilized in image-based learning when applied to WiFi CSI. We collect and make publicly available a dataset of CSI amplitude spectrograms of human activities. Utilizing this data, an ablation study is conducted in which we train activity recognition models based on the EfficientNetV2 architecture, allowing us to evaluate the impact of each augmentation on model generalization performance. The results show that, although no single technique is universally effective, specific combinations of data augmentations applied to CSI amplitude features can significantly enhance generalization in certain cross-scenario and cross-system settings. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.