From Clutter to Clarity: Enhancing Radar-Based Human Activity Recognition with Deep Attention and Feature Denoising

Millimeter-wave radar offers advantages such as insensitivity to lighting conditions, strong environmental adaptability, non-contact sensing, and inherent protection of user privacy. These characteristics have led to its increasing application in human activity recognition, with promising use cases...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE sensors journal S. 1
Hauptverfasser: Wang, Huaijun, Li, Shuang, Bai, Bingqian, Li, Junhuai, Fei, Rong, Huang, Tao
Format: Journal Article
Sprache:Englisch
Veröffentlicht: IEEE 19.11.2025
Schlagworte:
attention mechanism
Convolutional Autoencoder
human action recognition
millimeter wave radar
ISSN:1530-437X, 1558-1748
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Millimeter-wave radar offers advantages such as insensitivity to lighting conditions, strong environmental adaptability, non-contact sensing, and inherent protection of user privacy. These characteristics have led to its increasing application in human activity recognition, with promising use cases in smart homes, healthcare monitoring, and security systems. However, during radar data acquisition, substantial background interference and noise can significantly reduce recognition accuracy. To address this issue, this paper proposes a human action recognition method that integrates a deep convolutional autoencoder with a multi-scale attention mechanism. In the preprocessing stage, phase mean subtraction and moving average filtering are applied to suppress static background clutter in the radar micro-Doppler features. Subsequently, feature denoising and enhancement are performed using a convolutional autoencoder and attention modules to improve focus on motion-relevant regions. Experimental results demonstrate that the proposed method achieves over 96% accuracy on both a self-constructed dataset and a public benchmark dataset. Furthermore, the model maintains high recognition performance under low signal-to-noise ratio conditions, confirming its robustness and effectiveness in complex environments.
ISSN:1530-437X
1558-1748
DOI:10.1109/JSEN.2025.3633200