# Attention-Guided Track-Pulse-Sequence Target Association Network

## TL;DR

This paper introduces a new network for associating satellite tracks with maritime targets using attention mechanisms and pulse sequences, improving accuracy in dense tracking scenarios.

## Contribution

The novel AG-TPS-TAN network combines track and pulse-sequence data with an attention mechanism and dual-feature fusion for better target association.

## Key findings

- AG-TPS-TAN achieved 93.91% accuracy for 5 targets and 63.83% for 50 targets.
- Outperformed track-only and signal-statistics-based methods in dense target scenarios.

## Abstract

Multi-satellite sequential detection is crucial for maritime target identification and tracking. However, inherent satellite revisit patterns and maritime target motion often result in fragmented track segments, necessitating effective multi-satellite track association to ensure continuity. Existing methods predominantly rely on track information and statistical signal parameters, rendering them susceptible to localization errors and ineffective in scenarios characterized by dense targets and overlapping radar parameters. To overcome these limitations, this paper proposes an attention-guided track-pulse-sequence target association network (AG-TPS-TAN). First, the asymmetric dual-branch network operates by incorporating both track data and electromagnetic signal data, processing the latter in the form of raw pulse sequences instead of the conventional statistical parameters. Second, within the track branch, we enhance the feature representation by incorporating a novel track-point-aware attention mechanism which can autonomously identify and weight critical points indicative of motion continuity, such as interruption boundaries and maneuvering points. Third, we introduce a dual-feature fusion module optimized with a combined loss function, which pulls feature representations of the same target closer together while pushing apart those from different targets, thereby enhancing both feature consistency and discriminability. Experiments were conducted on a public AIS trajectory dataset, constructing a dataset containing both motion trajectories and electromagnetic signals. Evaluations under varying target numbers showed that the proposed AG-TPS-TAN achieved average association accuracies of 93.91% for 5 targets and 63.83% for 50 targets. Against this, the track-only method TSADCNN scored 76.08% and 25.64%, and the signal-statistics-based method scored 77.12% and 29.56%, for 5 and 50 targets, respectively, thus exhibiting a clear advantage for the proposed approach.