Ballistic target recognition based on multiple data representations and deep-learning algorithms

Target recognition is a significant part of a Ballistic Missile Defense System (BMDS). However, most existing ballistic target recognition methods overlook the impact of data representation on recognition outcomes. This paper focuses on systematically investigating the influences of three novel data representations in the Range-Doppler (RD) domain. Initially, the Radar Cross Section (RCS) and micro-Doppler (m-D) characteristics of a cone-shaped ballistic target are analyzed. Then, three different data representations are proposed: RD data, RD sequence tensor data, and RD trajectory data. To accommodate various data inputs, deep-learning models are designed, including a two-Dimensional Residual Dense Network (2D RDN), a three-Dimensional Residual Dense Network-Gated Recurrent Unit (3D RDN-GRU), and a Dynamic Trajectory Recognition Network (DTRN). Finally, an Electromagnetic (EM) computation dataset is collected to verify the performances of the networks. A broad range of experimental results demonstrates the effectiveness of the proposed framework. Moreover, several key parameters of the proposed networks and datasets are extensively studied in this research.