Transformer Based Multi-Target Bernoulli Tracking for Maritime Radar

TL;DR

This paper introduces a transformer-based method for multi-target maritime radar tracking, improving detection and tracking accuracy over traditional CFAR techniques by leveraging machine learning and advanced filtering.

Contribution

It presents a novel transformer-based measurement extraction and a measurement-driven birth density design for multi-target tracking in maritime radar, outperforming traditional methods.

Findings

Transformer approach outperforms CFAR in all scenarios

Measurement-driven birth density improves tracking accuracy

Superior error performance demonstrated in experiments

Abstract

Multi-target tracking in the maritime domain is a challenging problem due to the non-Gaussian and fluctuating characteristics of sea clutter. This article investigates the use of machine learning (ML) to the detection and tracking of low SIR targets in the maritime domain. The proposed method uses a transformer to extract point measurements from range-azimuth maps, before clustering and tracking using the Labelled mulit- Bernoulli (LMB) filter. A measurement driven birth density design based on the transformer attention maps is also developed. The error performance of the transformer based approach is presented and compared with a constant false alarm rate (CFAR) detection technique. The LMB filter is run in two scenarios, an ideal birth approach, and the measurement driven birth approach. Experiments indicate that the transformer based method has superior performance to the CFAR approach for all target scenarios discussed