Random Code Radar With Range-Time–Frequency Points and Improved PointConv Network for Through-Wall Human Action Recognition

We proposed and demonstrated experimentally a random code radar with range-time–frequency points and the improved PointConv network for through-wall human action recognition (HAR). The physical random code signal with the natural random and aperiodicity is used as the radar-transmitted waveform. A series of slow time-Doppler frequency (ST-DF) images are obtained by correlation ranging and short-time Fourier transform (STFT) for echo and reference signals, and then are arranged at different ranges to obtain the 3-D range-time–frequency matrix. The range-time–frequency points are input into the improved PointConv network after the constant false alarm rate (CFAR) detection, isosurface mesh generation (IMG), and farthest point sampling (FPS) for the 3-D matrix. The PointConv network is improved by model simplification and structural enhancement, which can achieve higher recognition accuracy, the smaller parameters with 5.53 M, and the smaller floating-point operations (FLOPs) of 1.06 G, compared to the existing PointConv network. Experimental results demonstrate that the proposed radar can accurately recognize human actions behind walls with a 99.63% average accuracy for ten actions and a 96.83% average accuracy for six participants. Compared with the 2-D image-based convolutional neural network, three-domain feature fusion, two 3-D point-based PointNet networks, and 3-D point-based PointConv network, the proposed method realizes the higher recognition accuracy.