In this paper, a multi-feature encoder for gesture recognition based on a 60 GHz frequency-modulated continuous wave (FMCW) radar system is proposed to extract the gesture characteristics, i.e., range, Doppler, azimuth and elevation, from the low-level raw data. The radar system updates the hand information for every measurement-cycle on all the scattering centers in its field of view, and our proposed encoder is devised to only focus on those essential scattering centers. After observing the hand over several measurement-cycles, we encode the gesture characteristics sequentially into a 2-D feature matrix, which is successively fed into a shallow convolutional neural network (CNN) for classification. For the purpose of distinguishing relevant gestures, the proposed multi-feature encoder is able to efficiently extract adequate information from a multi-dimensional feature space. Thus, the proposed approach is practical for industrial applications where the available dataset is mostly small-scale. The experimental results show that the proposed multi-feature encoder could guarantee a promising performance for a gesture dataset with 12 gestures.
展开▼
