Real-time rail vehicle localisation using spatially resolved magnetic field measurements

TL;DR

This paper introduces two real-time magnetic field-based rail vehicle localization methods, demonstrating high accuracy and robustness, especially in cold-start scenarios, suitable for safety-critical rail systems.

Contribution

It presents a novel hybrid localization system combining particle filtering and sequence alignment for improved accuracy and cold-start performance.

Findings

Particle filter achieves sub-5-meter accuracy over 21.6 km

Sequence alignment localizes within 30 m in 92% of cold-start tests

System runs in real-time on consumer hardware

Abstract

This work presents two complementary real-time rail vehicle localization methods based on magnetic field measurements and a pre-recorded magnetic map. The first uses a particle filter reweighted via magnetic similarity, employing a heavy-tailed non-Gaussian kernel for enhanced stability. The second is a stateless sequence alignment technique that transforms real-time magnetic signals into the spatial domain and matches them to the map using a similarity measure. Experiments with operational train data show that the particle filter achieves track-selective, sub-5-meter accuracy over 21.6 km, though its performance degrades at low speeds and during cold starts. Accuracy tests were constrained by the GNSS-based reference system. In contrast, the alignment-based method excels in cold-start scenarios, localizing within 30 m in 92 % of tests (100 % using top-3 matches). A hybrid approach combines both methods$\unicode{x2014}$alignment-based initialization followed by particle filter tracking. Runtime analysis confirms real-time capability on consumer-grade hardware. The system delivers accurate, robust localization suitable for safety-critical rail applications.