Time-Space Dimension Reduction of Millimeter-Wave Radar Point-Clouds for Smart-Home Hand-Gesture Recognition

Abstract

This paper proposes a novel time-space dimension reduction method for millimeter-wave radar point-clouds on long-distance or large-field-of-view hand-gesture recognition. First, a two-dimensional antenna array is used to obtain radar point-clouds composed of range, Doppler, azimuth, elevation, and amplitude. The point-cloud time series are further reduced in dimensionality to reconstruct radar spectrums with less interference and better time-space continuity in the Range-Doppler-Azimuth-Elevation (RDAE) domain, and generate new spatial position spectrums after transforming to the three-dimensional cartesian coordinate system, which are superior to radar spectrums in the RDAE domain in terms of space invariance. Then a spatial position alignment method is proposed to improve the spatial consistency of multi-position dataset. Finally, a multi-channel convolutional neural network is designed for multi-position hand-gesture recognition. Experimental results show that the proposed method can solve the problems of poor space invariance in the RDAE domain and poor generalization performance for different spatial positions and different users in existing methods. For hand-gesture recognition within a long distance of 3 meters and a large field-of-view of 60 degrees, the average classification accuracies of 5 waving hand-gestures collected at multiple positions for trained user A, non-trained user B, and non-trained user C can reach 99.7%, 91.1%, and 87.1%, respectively.