PALReg: Polarization-Aspect Fusion Lightweight Regressor for Micromotion Cone-Shaped Objects

Abstract

The Doppler effect is essential for confirming aerial cone-shaped objects. However, individual radars are limited in the detection and analysis capabilities due to the viewing angles and polarization. PALReg, a deep learning (DL)-based approach, leverages electromagnetic (EM) scattering information from multiple radar angles and polarizations to deduce the micromotion and geometric parameters of aerial cone-shaped objects with Doppler effect. A custom dual-branch backbone network is employed, enhancing feature diversity and model robustness by capturing more complementary information from spectrograms. The backbone incorporates both lightweight blocks and attention blocks, markedly reducing computational load while improving prediction accuracy. To optimize performance, a set of tailored weight coefficients are integrated into the loss function, targeting specific predicted physical parameter and further enhancing effectiveness of the regressor. To validate the efficacy of PALReg, a dataset comprising three types of cone-shaped objects, each under varying micromotion states and geometric configurations is constructed. Experimental results show that PALReg surpasses the existing DL-based models, achieving high accuracies of 98.86%, 98.13%, 98.86%, 98.39%, and 98.56% across five parameters, with a model size under 7.50 MB.