Domain-Independent Gesture Recognition Using Single-Channel Time-Modulated Array

• Published in: IEEE transactions on antennas and propagation Vol. 72; no. 4; pp. 3386 - 3399
• Main Authors: Guan, Lei, Yang, Xiaodong, Zhao, Nan, Alomainy, Akram, Imran, Muhammad Ali, Abbasi, Qammer H.
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Doppler radar; Fast Fourier transformations; Feature extraction; Fourier transforms; Gesture recognition; Modulation; Moving targets; neural network; Neural networks; Radar; Radio frequency; Sensors; time-modulation array (TMA)
• ISSN: 0018-926X; 1558-2221
• DOI: 10.1109/TAP.2024.3373054

In recent years, gesture recognition system based on radio frequency (RF) sensing has a wide application prospect and attraction in noncontact electronic interaction with its advantages of privacy security, lighting independence, and wide sensing range. The traditional RF sensing system depends on the environment and the subject, and the multichannel sensing equipment is expensive, which brings great challenges to the practical application. To address the above issues, a single-channel, low-cost, and domain-independent gesture recognition system is proposed. Specifically, the time-modulation technology is adopted to expand the number of antennas of the sensing device. The time-modulation array (TMA) is converted into a traditional array through harmonic recovery technology. The 2D-fast Fourier transform (FFT), moving target indication filter, and data normalization are used to extract domain-independent angle-Doppler maps (ADMs) gesture features. In order to ensure recognition accuracy, we propose a lightweight neural network with an attention mechanism, which only needs one training and can be applied to different data domains. The experimental results show that the accuracy of in-domain recognition of the proposed system is 98.9%, and the accuracy of cross-domain (i.e., new environments, new users, and new positions) recognition is 85.6%-97.4% without model retraining.