Scholarlink

In this work, the problem of cross-environment generalization in WiFi Channel State Information (CSI)-based localization and Human Activity Recognition (HAR) models within through-wall scenarios is addressed, highlighting an area that remains underexplored. A comprehensive evaluation is conducted to investigate the effectiveness of various methodologies, including CSI feature selection, feature scaling, dimensionality reduction, and data augmentation techniques, in improving model robustness to environmental variations. The evaluation is based on a dataset collected over three days in environments exhibiting both static and dynamic variations, featuring synchronized CSI and 3D trajectory data of human activities, which is made publicly available at https://zenodo.org/records/10925351 . The findings reveal that plain CSI amplitude features consistently outperform other types in achieving superior generalization in through-wall scenarios. Furthermore, it is found that while dimensionality reduction techniques like PCA, ICA, and UMAP do not enhance model generalization, feature scaling and data augmentation can significantly improve both localization and HAR performance in the presence of static and dynamic environmental variations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

On the Generalization of WiFi-Based Person-Centric Sensing in Through-Wall Scenarios

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Julian Strohmayer,
  • Martin Kampel

摘要

In this work, the problem of cross-environment generalization in WiFi Channel State Information (CSI)-based localization and Human Activity Recognition (HAR) models within through-wall scenarios is addressed, highlighting an area that remains underexplored. A comprehensive evaluation is conducted to investigate the effectiveness of various methodologies, including CSI feature selection, feature scaling, dimensionality reduction, and data augmentation techniques, in improving model robustness to environmental variations. The evaluation is based on a dataset collected over three days in environments exhibiting both static and dynamic variations, featuring synchronized CSI and 3D trajectory data of human activities, which is made publicly available at https://zenodo.org/records/10925351 . The findings reveal that plain CSI amplitude features consistently outperform other types in achieving superior generalization in through-wall scenarios. Furthermore, it is found that while dimensionality reduction techniques like PCA, ICA, and UMAP do not enhance model generalization, feature scaling and data augmentation can significantly improve both localization and HAR performance in the presence of static and dynamic environmental variations.