Experimental Evaluation of Monocular Visual–Inertial SLAM Methods for Freight Railways

Accurate, real-time, seamless, and reliable train localization is one of the most critical tasks for automatic train control systems. With the prospect of replacing the traditional track-side infrastructures, independent on-board sensors have attracted a great deal of interest. Since each rail vehicle is equipped with visual cameras, the visual–inertial simultaneous localization and mapping (VI-SLAM) is a promising approach for train localization due to its simple sensor setups. However, the existing VI-SLAM methods mostly pay attention to indoors, urban, or suburban environments, where freight railways in the industrial scenarios have not been considered or investigated overall. In this article, we seek to explore the potential of VI-SLAM for railways, and provide a comprehensive experimental evaluation of several typical monocular VI-SLAM methods for freight railways. Our datasets are collected on the Chinese Fuyang–Luan railway utilizing a maintenance rail vehicle. The results illustrate that monocular VI-SLAM methods are extremely easy to fail, even within a few minutes. These methods also need to deal with severe scale estimation errors due to constrained motions.

I. Introduction

The deployment of Chinese railway networks has expanded significantly in the past years. By the end of 2021, the length of railways in operation has reached150 000 km, including more than 40000 km of high-speed railways. In order to keep up with the sustainable growth, the Chinese railway government proposes a superior level of requirements and standards for the current railway train control system [1]. With the aim of achieving accurate, safe, reliable, and real-time application, train localization is performing an increasingly crucial role in the railway transportation system [2].