Machine Learning Models for Improved Tracking from Range-Doppler Map Images

TL;DR

This work tackles the challenge of unreliable uncertainty propagation in ML-driven GMTI radar tracking by introducing a coupled pipeline: a UNet-based target detector for RDM images and a CVAE-based uncertainty estimator conditioned on the detector’s features. The authors demonstrate that integrating these models into a multi-hypothesis tracker yields significant improvements in tracking accuracy and localization over traditional STAP+CFAR baselines, particularly in complex multi-target scenarios. Key contributions include a discriminative UNet detector tailored to single-pixel-like targets amidst clutter, a conditional CVAE to model residual uncertainty, and comprehensive experiments using a synthetic but realistic sensor model. The approach has practical impact for GMTI tracking by providing both higher detection performance and principled uncertainty estimates that enhance downstream estimation chained through MHT, with feasible computational costs on modern hardware.

Abstract

Statistical tracking filters depend on accurate target measurements and uncertainty estimates for good tracking performance. In this work, we propose novel machine learning models for target detection and uncertainty estimation in range-Doppler map (RDM) images for Ground Moving Target Indicator (GMTI) radars. We show that by using the outputs of these models, we can significantly improve the performance of a multiple hypothesis tracker for complex multi-target air-to-ground tracking scenarios.

Figures (10)

• Figure 1: The red lines indicate the endo-exo divide separating the endo-clutter region inside the lines with the exo-clutter region outside the lines.
• Figure 2: Example of a UNet architecture: Blue boxes are multi-channel feature maps, white boxes are copied feature maps, and the arrows denote the different operations.
• Figure 3: Example of the target response (dark red center pixel) in a RDM image bleeding into its neighboring pixel (bright red) due to the discretization from binning of the target response.
• Figure 4: The architecture of one of the twin (endo- and exo-clutter) CVAEs in the statistical model of the target detector.
• Figure 5: Receiver Operating Characteristic (ROC) curve: Red is the traditional STAP prepossessing followed by the CFAR test, Blue is our UNet based target detector. Inner figure is plotted with the false positive rate (FPR) on a log scale.