Time-Varying Parametric Scattering Model Guided Network for Multidimensional Parameters Estimation of Dynamic Group Cone-Shaped Targets

Abstract

This article proposes a time-varying parametric scattering model guided network that estimates the multidimensional parameters of group cone-shaped targets. First, the time-varying parametric scattering model is extracted by an inverse parametric modeling method, which eliminates the false scattering centers (SCs), supplements the occluded SCs, and separates the overlapped scattering models of the dynamic group cone-shaped targets. The multiple scattering features of isolated targets are then reconstructed by the time-varying parametric scattering model. Second, a scattering model guided network that extracts the deep learning and physical features from the scattering features is proposed. The network is trained under physics-guided loss functions based on the time-varying ranges and micro-Doppler shifts of the time-varying parametric scattering model. Third, a cross-feature fusion block that adaptively fuses deep learning and physical features is designed. The target parameters are estimated from the fusion features through a fully connected (FC) layer. Finally, the effectiveness of the proposed method is evaluated through a simulation analysis.