Scholarlink

<p>In the field of Human Activity Recognition (HAR) with WiFi, the temporal variability of wireless channels results in the inability of pre-trained models in the source domains to adapt to unknown target domains. Existing cross-domain HAR works focus on handling domain shifts caused by domain factors to enhance the adaptability of models. Nonetheless, these works consider different domains as independent factors causing domain shifts, ignoring the correlation between domain factors. In this regard, we propose a cross-domain HAR system with Multi-domain factors Disentanglement and Test time Adaptation (MDTA). MDTA designs a feature extraction method based on domain factors disentanglement which uses nonlinear dimensionality augmentation mapping to disentangle domain correlations, and subsequently reconstructs the independent features of each domain factor through feature segmentation. The system further separates irrelevant factors in activity features based on triplet loss to bridge domain shift. Additionally, MDTA introduces an activity features classification method based on KNN-Graph classifier, it replaced the traditional KNN classifier’s voting mechanism based on local feature similarity with a global prediction confidence transfer method based on graph networks, and improve the adaptability of classifier to the feature distribution of target domain by updating the stored activity features at test time. The proposed MDTA system is evaluated using the Widar3.0 dataset and achieves recognition accuracies of over 90% across four cross-domain tasks. Code are available at <a href="https://github.com/jgkwo/MDTA">https://github.com/jgkwo/MDTA</a>.</p>

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

MDTA: CSI based cross-domain human activity recognition with multi-domain factors disentanglement and test time adaptation

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Yong Zhang,
  • Guangkun Ji,
  • Weiwei Jiang,
  • Changchuang Zhao,
  • Yujie Wang,
  • Lingyue Sun

摘要

In the field of Human Activity Recognition (HAR) with WiFi, the temporal variability of wireless channels results in the inability of pre-trained models in the source domains to adapt to unknown target domains. Existing cross-domain HAR works focus on handling domain shifts caused by domain factors to enhance the adaptability of models. Nonetheless, these works consider different domains as independent factors causing domain shifts, ignoring the correlation between domain factors. In this regard, we propose a cross-domain HAR system with Multi-domain factors Disentanglement and Test time Adaptation (MDTA). MDTA designs a feature extraction method based on domain factors disentanglement which uses nonlinear dimensionality augmentation mapping to disentangle domain correlations, and subsequently reconstructs the independent features of each domain factor through feature segmentation. The system further separates irrelevant factors in activity features based on triplet loss to bridge domain shift. Additionally, MDTA introduces an activity features classification method based on KNN-Graph classifier, it replaced the traditional KNN classifier’s voting mechanism based on local feature similarity with a global prediction confidence transfer method based on graph networks, and improve the adaptability of classifier to the feature distribution of target domain by updating the stored activity features at test time. The proposed MDTA system is evaluated using the Widar3.0 dataset and achieves recognition accuracies of over 90% across four cross-domain tasks. Code are available at https://github.com/jgkwo/MDTA.