WiFi-Enabled Gesture Recognition Using Attention-enhanced DenseNet

The widespread application of WiFi sensing technology has received a lot of research attention. But when the same person makes the same gesture in different orientations, or different people make the same posture in the same environment, it can lead to significant differences in the received WiFi si...

Full description

Saved in:
Bibliographic Details
Published in2024 IEEE/CIC International Conference on Communications in China (ICCC) pp. 1692 - 1697
Main Authors Yu, Xinlong, Li, Baogang, Chen, Jiale
Format Conference Proceeding
LanguageEnglish
Published IEEE 07.08.2024
Subjects
Accuracy
channel state information
Computational modeling
Coord-Attention
Data preprocessing
DenseNet
Gesture recognition
Image recognition
Noise
Sensors
WiFi sensing
Online AccessGet full text

Cover

More Information
Summary:The widespread application of WiFi sensing technology has received a lot of research attention. But when the same person makes the same gesture in different orientations, or different people make the same posture in the same environment, it can lead to significant differences in the received WiFi signal, which is the problem of cross domain recognition of signals. Therefore, this paper proposes a gesture recognition system that can realize cross-user and cross-orientation. The system consists of two modules: data preprocessing and gesture recognition. In the data preprocessing module, we remove the noise from the channel state information, and propose a dynamic similar subcarrier selection method to obtain high-quality subcarriers. The obtained data will then be further converted into RGB images. In the gesture recognition module, we use DenseNet as the backbone network in conjunction with an Coord-Attention module to achieve RGB image recognition. The accuracy of the system in in-domain gesture recognition is 98.89\%. The recognition accuracy across-users and across-orientations is 98.42\% and 97.66\%, respectively, which is better than existing recognition methods.
DOI:10.1109/ICCC62479.2024.10681945