Simulation of Deep Learning-Based Multitarget Track Association for Ballistic Target Groups

Abstract

This article focuses on the midcourse track association scenario of ballistic target groups (BTGs) observed by ground-based pulse-Doppler radar. It proposes a BTG track association neural network (BTGTANN) to perform track detection and association for individual targets within a BTG. First, time–range profile (TRP) samples generated by performing pulse compression (PC) on raw echo signals are used to represent the spatial distribution of multiple targets over time. Second, a feature selection and aggregation (FSA) module and a context-aware enhancement (CAE) module are developed based on a convolutional neural network (CNN) architecture. These modules enhance the feature fusion and context awareness capabilities of the network. Finally, the target detection branch of the BTGTANN is used to detect multiple target tracks in TRP samples, yielding track detection boxes. An instance segmentation branch is then employed to accurately extract the contours of the tracks within the detection boxes, thereby determining the track positions at each pulse time. Unlike traditional methods, this approach formulates the multitarget track association problem as an object detection and instance segmentation task, providing an innovative solution within a deep learning framework. Experimental results on simulated datasets demonstrate that the detection probability ( ${P}_{d}$ ), the false alarm probability ( ${P}_{f}$ ), and the root-mean-square error (RMSE) of the BTGTANN reached 93.81%, 0.11%, and 8.43 m, respectively. Relative to the baseline, ${P}_{d}$ was increased by 5.70%, while ${P}_{f}$ and RMSE were decreased by 0.06% and 3.97 m, respectively. Moreover, the robustness of the BTGTANN is validated across different target scenarios, with the results indicating its substantial performance and generalizability under multiple targets, low-signal-to-noise ratio (SNR), and low-signal-to-clutter ratio (SCR) environments.