A System for Image-Based Non-Line-Of-Sight Detection Using Convolutional Neural Networks

Abstract

The ERSAT GGC project introduces the concept of virtual balises for train localization, which avoids investment and maintenance costs of physical balises. Since this concept relies on the matching of train positions to balise positions stored in a database, it is dependent on placing virtual balises in track areas with unimpeded GNSS reception. One factor majorly contributing to the distortion of GNSS signals is the non-line-of-sight (NLOS) scenario where the direct path between a satellite and the receiver on the train is blocked. As these NLOS situations result in deflections or the total absence of GNSS signals, this paper proposes a system to identify obstacles occluding the visibility of satellites above the tracks traversed by a train. This is achieved by video recording the sky from the roof of the train and segmenting the images into sky and non-sky regions. The line-of-sight status of individual satellites is found through projecting the known satellite locations into the segmented images. Consequently, the information whether a satellite is located in a sky or non-sky segment of the image allows for a determination of the GNSS performance at any observed track area.