RRN-ATR-Net: A Robust Relation Network for Aerial Targets Recognition Based on Two-Stream Micro-Doppler Signatures Fusion

Unmanned aerial vehicles (UAVs) have become a challenge for radar to monitor aerial targets due to the “low, slow, and small” characteristics. In a recent study, micro-Doppler signatures (m-Ds) have shown the potential in distinguishing aerial targets. However, collecting large amounts of radar data containing m-Ds for aerial targets’ recognition network training is time-consuming and laborious. Besides, the m-Ds displayed by the time-Doppler map (TDM) is prone to deform with the measurement angle. To solve these problems, a robust relation network for aerial targets’ recognition (RRN-ATR-Net) is proposed. First, a few-shot network requiring fewer samples is introduced to determine the type of aerial targets through similarity with the supporting samples. Second, time-Cepstrum map (TCM), which is independent of measurement, is adopted as the complementary signature of the TDM, and a tensor outer product-based micro-Doppler (m-D) fusion module is proposed to capture the mutual features for robustness improvement. Third, a transformer-based relation module is proposed to construct long-range correlation dependencies to generate attention maps with nonlocal and high semantic contrast that can guide the network to learn key information. Multiangle and multidistance measurements are performed in a field environment. Experimental results show that when there are only five, seven, and ten supporting samples in each class, the proposed network achieves the accuracies of 92.65%, 95.85%, and 96.55%, respectively.