Semantic Segmentation Guided RF Micro-Doppler Synthesis and UAV Classification in Low SINR

This paper presents a novel approach to synthesize and classify UAV radar micro-Doppler (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \mathrm{D}$</tex>) signatures by utilizing semantic segmentation. The micro-Doppler signatures of UAVs are characterized by Helicopter Motor Rotation (HERM) lines, which appear in parallel to the main radar return and contain information relating to the number of and rotation rate of propellers of the UAV. The rich information in the HERM lines make them of importance when not only classifying but also synthesizing UAV <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex> D signatures. Existing radar <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu \mathrm{D}$</tex> signature synthesis methods do not explicitly account for the inherent physics of HERM lines, and therefore result in kinematically inaccurate synthetic samples. We propose a novel method that preserves the kinematics of HERM lines by integrating semantic segmentation into the discriminator of a Generative Adversarial Network (GAN). Low-fidelity physics-based simulations of the radar return from UAVs are utilized to pre-train U-Net, thereby enabling the semantic segmentation of HERM lines across a variety of environments with varying noise/clutter levels. Our results show that UAV <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mu$</tex> D signatures synthesized using the proposed Semantic Segmentation GAN (SS-GAN) enabled improved model training, even in low SINR environments. Furthermore, we propose SS-DNN, which exploits semantic segmentation pretrained with physics-based simulations as a signal conditioner prior to input to a deep neural network, for classification. Results show that the combination of SS-GAN with SS-DNN offers an improvement of 8 % for UAV recognition in low SINR.