Scholarlink

<p>Recent technological advancements have spurred a surge in diverse wearable devices. Many of them incorporate inertial sensors, enabling fundamentally new methodologies in personal recognition based on gait patterns. However, previous research tends to focus on specific walking scenarios and sensor placements, limiting applicability in diverse real-world contexts. This study presents a novel neural network framework that efficiently harnesses the strengths of Convolutional Neural Networks and Multi-Head Attention to enable robust gait-based identity recognition, demonstrating adaptability across multiple scenarios. We evaluated it on two public inertial datasets (the OU-ISIR Inertial Sensor Dataset; and whuGAIT - Dataset 3) and a custom inertial dataset (CDUT-IG). For each dataset, a distinct combination of wearable devices and on-body placement locations was employed to collect gait data. Our method achieved recognition accuracy of 95.55%, 98.58%, and 99.69% on the whuGAIT, OU-ISIR Inertial Sensor Dataset, and CDUT-IG, respectively. Results validate the robustness of the proposed method for identity recognition using inertial gait data collected through various types of wearable devices. In addition, we employed the Integrated Gradient algorithm to investigate the specific contribution of individual regions within the input data series to the classification result made by the models.</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

Advancing wearable gait-based identity recognition: multi-head attention and feature analysis for unconstrained environments and multiple sensor locations

  • 期刊论文

  • 发表日期：

    2026年2月14日

查看全文
  • Tong He,
  • Sijia Yi,
  • Hui Zeng,
  • Zijie Mei,
  • Kamen Ivanov,
  • Guocheng Yang,
  • Lin Yang,
  • Jie Tang,
  • Zhanyong Mei

摘要

Recent technological advancements have spurred a surge in diverse wearable devices. Many of them incorporate inertial sensors, enabling fundamentally new methodologies in personal recognition based on gait patterns. However, previous research tends to focus on specific walking scenarios and sensor placements, limiting applicability in diverse real-world contexts. This study presents a novel neural network framework that efficiently harnesses the strengths of Convolutional Neural Networks and Multi-Head Attention to enable robust gait-based identity recognition, demonstrating adaptability across multiple scenarios. We evaluated it on two public inertial datasets (the OU-ISIR Inertial Sensor Dataset; and whuGAIT - Dataset 3) and a custom inertial dataset (CDUT-IG). For each dataset, a distinct combination of wearable devices and on-body placement locations was employed to collect gait data. Our method achieved recognition accuracy of 95.55%, 98.58%, and 99.69% on the whuGAIT, OU-ISIR Inertial Sensor Dataset, and CDUT-IG, respectively. Results validate the robustness of the proposed method for identity recognition using inertial gait data collected through various types of wearable devices. In addition, we employed the Integrated Gradient algorithm to investigate the specific contribution of individual regions within the input data series to the classification result made by the models.