Scholarlink

<p>In complex indoor environments, Wi-Fi signals are susceptible to obstruction and multi-path effects, limiting the performance of traditional positioning methods. To address this issue, this paper proposes a Wi-Fi indoor localization model that integrates multitask learning with spatiotemporal features. Using a CNN–GRU architecture, the model extracts deep spatial and temporal information and jointly optimizes building identification, floor classification, and coordinate regression to enhance generalization and discriminative capability. Channel attention, a generative adversarial network, and a denoising autoencoder are further incorporated to improve feature representation and data robustness. Experiments on the UJIIndoorLoc dataset demonstrate that the proposed method outperforms traditional approaches and several recent deep learning models in terms of building identification accuracy (99.82%), floor classification accuracy (91.89%), and average localization error (9.38&#xa0;m). Owing to its computation-intensive nature, the model is efficiently trained and evaluated using high-performance computing (HPC) resources. The code of our research work is publicly available at: <a href="https://github.com/JK093878/gitt">https://github.com/JK093878/gitt</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

Wi-fi indoor positioning algorithm based on multitask spatiotemporal fusion

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Shuang Zhai,
  • Yanzhao Qiu,
  • Xiao Zhao,
  • Zihao Lu,
  • Yongqi Lyu

摘要

In complex indoor environments, Wi-Fi signals are susceptible to obstruction and multi-path effects, limiting the performance of traditional positioning methods. To address this issue, this paper proposes a Wi-Fi indoor localization model that integrates multitask learning with spatiotemporal features. Using a CNN–GRU architecture, the model extracts deep spatial and temporal information and jointly optimizes building identification, floor classification, and coordinate regression to enhance generalization and discriminative capability. Channel attention, a generative adversarial network, and a denoising autoencoder are further incorporated to improve feature representation and data robustness. Experiments on the UJIIndoorLoc dataset demonstrate that the proposed method outperforms traditional approaches and several recent deep learning models in terms of building identification accuracy (99.82%), floor classification accuracy (91.89%), and average localization error (9.38 m). Owing to its computation-intensive nature, the model is efficiently trained and evaluated using high-performance computing (HPC) resources. The code of our research work is publicly available at: https://github.com/JK093878/gitt.