Through-Wall Human Activity Recognition With Complex-Valued Range–Time–Doppler Feature and Region-Vectorization ConvGRU
Longzhen Tang, Shisheng Guo, Jian Qiang, Guolong Cui, Lingjiang Kong, Xiaobo Yang
Abstract
In this paper, we consider a high-accuracy and low-complexity method for recognizing human activities behind wall. As the amount of information conveyed by data representation directly affects the recognition accuracy of the network, we construct the three-dimensional (3D) complex-valued feature for human activity recognition (HAR). In light of high network complexity introduced by 3D complex-valued data, we devise a low time and space complexity network named Convolutional Gated Recurrent Unit based on region-vectorization (RV-ConvGRU). Keystone Transform is utilized to process the radar echo and generate 3D complex-valued Range-Time-Doppler (RTD) data first, which provides high-frequency resolution and abundant feature information. Then, the real and imaginary parts of the complex-valued RTD are separately fed into a feature extraction module to comprehensively extract their respective features. Specifically, the real or imaginary part of the RTD is divided into multiple regions, which are then converted into regional vectors and reordered as channels to reduce the time and space complexity of the subsequent network. The reconfigured features are then input into the Convolutional Gated Recurrent Unit (ConvGRU) to extract global and temporal features, with the channel attention mechanism for feature selecting. The features of the real and imaginary parts are fused and then classified by the classifier finally. The experiments verify that the proposed method is effective, achieving the highest recognition accuracy of 99.23% with an input sequence of 1.44 seconds.