EUVIMEDEUVIMED entdecken
Crossref

Open-Set Human Activity Recognition With Micro-Doppler Signatures and Virtual Prototype Learning

Kuiyu Chen, Chen Liu, Yunchao Song, Lingzhi Zhu

IEEE Transactions on Radar Systems · 2025 · 3 · 1463-1473

DOI: 10.1109/trs.2025.3628294 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Radar Systems, 2025. Verfasst von Kuiyu Chen und 3 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 41 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Jiangsu Province, Foundation of Key Laboratory of Radar Imaging and Microwave Photonics.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Human activity recognition (HAR) has emerged as a key technology, with applications ranging from security to healthcare. Radar-based HAR, which leverages micro-Doppler signatures, offers strong performance in complex environments. However, most existing systems operate under closed-set assumptions, recognizing only predefined activities. This restricts their effectiveness in real-world scenarios where novel or unseen activities may occur. To address this challenge, this work proposes a virtual prototype learning (VPL) framework for open-set HAR. Inspired by human memory and pattern-matching processes, VPL encodes micro-Doppler spectrograms into abstract representations and compares them with learned prototypes in the metric space. The framework is guided by a combination of Euclidean cross-entropy loss and clustering loss to promote clear separation between different activity classes while preserving consistency within each class. To further improve robustness, VPL incorporates a manifold mixup strategy, generating pseudo-samples that help sharpen the boundary between known and unknown activities. A buffer zone is established in the feature space to reinforce this separation, and hyperspherical decision boundaries are employed to enhance classification accuracy. Experiments with real-world radar data show that VPL outperforms existing methods, achieving higher accuracy for known activities while effectively detecting unknown activities.

Förderung

  • National Natural Science Foundation of China
  • Natural Science Foundation of Jiangsu Province
  • Foundation of Key Laboratory of Radar Imaging and Microwave Photonics
  • Natural Science Research Start-Up Foundation of Recruiting Talents of Nanjing University of Posts and Telecommunications

Zitieren

APA

Chen, K., Liu, C., Song, Y., & Zhu, L. (2025). Open-Set Human Activity Recognition With Micro-Doppler Signatures and Virtual Prototype Learning. IEEE Transactions on Radar Systems, 3, 1463-1473. https://doi.org/10.1109/trs.2025.3628294

Vancouver

Chen K, Liu C, Song Y, Zhu L. Open-Set Human Activity Recognition With Micro-Doppler Signatures and Virtual Prototype Learning. IEEE Transactions on Radar Systems. 2025;3:1463-1473. doi:10.1109/trs.2025.3628294

BibTeX
@article{Chen2025,
  author = {Chen, Kuiyu and Liu, Chen and Song, Yunchao and Zhu, Lingzhi},
  title = {Open-Set Human Activity Recognition With Micro-Doppler Signatures and Virtual Prototype Learning},
  journal = {IEEE Transactions on Radar Systems},
  year = {2025},
  volume = {3},
  pages = {1463--1473},
  doi = {10.1109/trs.2025.3628294}
}
Kostenlos starten →