Highly Accurate Video-Based Train Localization - replacing Balises with Natural Reference Points

Abstract

Recently, visual odometry has been successfully applied as a video-based approach across domains. We adapted this approach to railways achieving excellent results without using any other conventional rail sensors. Herewith, we propose an extension to our visual rail odometry approach that allows to visually compensate for the inevitable odometry drifts based on sporadically visible local scene structures and that provides means for a highly accurate train localization based on existing geo-referenced infrastructure of the rail system. The specific conditions of the visual rail navigation require an adaptation of the conventional VSLAM (Video-based Simultaneous Localization and Mapping) systems to cope with the limited and self-similar property of the observed area. We show how this extension can be used to replace the currently used train report system with a significantly increased global accuracy and reduced drift in the estimation between the geo-referenced rail structures like balises. Furthermore, a migration scenario is proposed which overcomes the issue of the approval of new localization systems. Area: Rail Navigation