Scholarlink

<p>WiFi-based gesture recognition has been extensively explored due to its characteristics of non-contact, non-line-of-sight and cost-effectiveness. However, it has overlooked the verification of the identity of the gesture performer, lacking in security measures. In this paper, we introduce DBN-IGR, a dual-branch network specifically designed to address the challenge by simultaneously handling identification and gesture recognition tasks. This architecture consists of a shared network and task-specific sub-networks, the former extracts common features relevant to both tasks, providing cross-domain capabilities, while the latter further extracts unique features tailored to each task. Notably, the entire network employs parallel stacked activation functions, which enhance its non-linearity by adjusting inputs from neighboring units to learn global information, thereby improving performance. The experimental results on the publicly available Widar3.0 dataset demonstrate excellent performance of our system compared to state-of-the-art systems, with average accuracies of 98.96% for identifying 9 individuals and 98.42% for recognizing 9 types of gestures. Moreover, in cross-domain scenarios, the average accuracies reach 89.087% and 90.364% respectively.</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

Dbn-igr: dual-branch network for identification and gesture recognition with WiFi signals

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Zhihua Li,
  • Shuli Ning,
  • Zhongcheng Wei,
  • Jijun Zhao

摘要

WiFi-based gesture recognition has been extensively explored due to its characteristics of non-contact, non-line-of-sight and cost-effectiveness. However, it has overlooked the verification of the identity of the gesture performer, lacking in security measures. In this paper, we introduce DBN-IGR, a dual-branch network specifically designed to address the challenge by simultaneously handling identification and gesture recognition tasks. This architecture consists of a shared network and task-specific sub-networks, the former extracts common features relevant to both tasks, providing cross-domain capabilities, while the latter further extracts unique features tailored to each task. Notably, the entire network employs parallel stacked activation functions, which enhance its non-linearity by adjusting inputs from neighboring units to learn global information, thereby improving performance. The experimental results on the publicly available Widar3.0 dataset demonstrate excellent performance of our system compared to state-of-the-art systems, with average accuracies of 98.96% for identifying 9 individuals and 98.42% for recognizing 9 types of gestures. Moreover, in cross-domain scenarios, the average accuracies reach 89.087% and 90.364% respectively.