EUVIMEDEUVIMED entdecken
Crossref

LPBS-Net: A Lightweight Network for Human Activity Recognition From Sparse Millimeter-Wave Radar Point Clouds

Fangfang Zhang, Hao Sun, Jinzhu Peng, Haijing Wang

IEEE Sensors Letters · 2025-12 · 9 (12) · 1-4

DOI: 10.1109/lsens.2025.3626751 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Sensors Letters, 2025-12. Verfasst von Fangfang Zhang und 3 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 15 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Millimeter-wave radar is widely used for indoor human activity recognition due to its privacy-preserving nature, with point cloud data effectively capturing target geometry. However, the sparsity and dynamic nature of these point clouds leads to unstable feature extraction, and resource constraints challenge large-scale neural network deployment. To address this, this letter proposes the Lightweight PointNet-BiLSTM with SE-Net (LPBS-Net), a lightweight network integrating a squeeze-and-excitation (SE) attention mechanism and bidirectional long short-term memory (BiLSTM) into a streamlined PointNet backbone, enhancing spatiotemporal feature modeling for dynamic point clouds. To overcome PointNet's need for fixed input point counts and its sensitivity to sparse distributions, we introduce Gaussian-based intensity and repeat padding, which selects base points by reflection intensity and uses Gaussian perturbation and repeated sampling to mitigate sparsity-induced feature degradation. Experiments on two public datasets show that LPBS-Net achieves 97.11% accuracy on the MMActivity dataset with only 0.176 M parameters, reducing model size by 84% compared to PointNet-BiLSTM, and outperforming existing methods, with maximum accuracy improvements exceeding 30%. The proposed lightweight network offers high accuracy and computational efficiency, evidenced by its low parameter count and floating point operations (FLOPs), making it suitable for deployment on resource-constrained edge devices.

Zitieren

APA

Zhang, F., Sun, H., Peng, J., & Wang, H. (2025). LPBS-Net: A Lightweight Network for Human Activity Recognition From Sparse Millimeter-Wave Radar Point Clouds. IEEE Sensors Letters, 9(12), 1-4. https://doi.org/10.1109/lsens.2025.3626751

Vancouver

Zhang F, Sun H, Peng J, Wang H. LPBS-Net: A Lightweight Network for Human Activity Recognition From Sparse Millimeter-Wave Radar Point Clouds. IEEE Sensors Letters. 2025;9(12):1-4. doi:10.1109/lsens.2025.3626751

BibTeX
@article{Zhang2025,
  author = {Zhang, Fangfang and Sun, Hao and Peng, Jinzhu and Wang, Haijing},
  title = {LPBS-Net: A Lightweight Network for Human Activity Recognition From Sparse Millimeter-Wave Radar Point Clouds},
  journal = {IEEE Sensors Letters},
  year = {2025},
  volume = {9},
  number = {12},
  pages = {1--4},
  doi = {10.1109/lsens.2025.3626751}
}
Kostenlos starten →