Low-Resolution Radar Target Classification Using Vision Transformer Based on Micro-Doppler Signatures
Beili Ma, Karen Egiazarian, Baixiao Chen
Abstract
Micro-Doppler signatures (m-DSs) have been widely employed for the automatic recognition of various radar targets that exhibit micromotions via time–frequency distributions (TFDs). However, most existing studies using time–frequency analysis for a good classification performance often require a continuous and long observation time to show stable and regular micromotion characteristics. In this article, we propose a single-frame recognition scheme based on a two-channel vision transformer (ViT) for low-resolution radar target classification. The proposed approach is achieved through three successive steps: one-frame radar signal generation, feature image representation, and a two-channel ViT network. In the first step, a one-frame radar signal for each coherent processing interval is generated based on a low-resolution pulsed radar system. Then, the short-time Fourier transform (STFT) and bispectrum are considered to fully excavate the m-DSs in the second step, and the energy- and phase-based feature images are represented in one-frame time. In the last step, we investigate a two-channel ViT network to realize the single-frame decision recognition. The effectiveness of the proposed two-channel ViT model, which fuses STFT and bispectrum features, is validated by the experimental results obtained from a group of measured radar data.