Dataset to &quot;GERALD: A novel dataset for the detection of German mainline railway signals&quot;

Leibner, Philipp Simon

doi:10.18154/RWTH-2024-02040

Dataset to "GERALD: A novel dataset for the detection of German mainline railway signals"

Leibner, Philipp Simon^RWTH*

2024

Online
DOI: 10.18154/RWTH-2024-02040
URL: https://publications.rwth-aachen.de/record/980030/files/GERALD.zip

Einrichtungen

Lehrstuhl und Institut für Schienenfahrzeuge und Transportsysteme (414210)

Inhaltliche Beschreibung (Schlagwörter)
railways (frei) ; advanced driver assistance systems (frei) ; object detection (frei) ; artificial intelligence (frei) ; autonomous trains (frei) ; computer vision (frei) ; automatic train protection (frei) ; railway signalling (frei)

Kurzfassung
In recent years, a strong push towards driverless mobility solutions can beseen in many transportation sectors including railways. While the EuropeanTrain Control System already specifies the necessary interfaces to open upthe possibility of Automatic Train Operation (ATO) for mainline railwayvehicles, required infrastructure-side upgrades of interlocking systems aretime- and cost-intensive. Alternatively, a pure vehicle-side Automatic TrainOperation solution can be conceptualized that relies on processing the sameaudio-visual input a human train driver would normally base his decisionson. This would require the vehicle-side detection of track-side railwaysignals to determine the vehicle’s movement authority and allowed maximumspeed. Such a signal detection system could furthermore be employed as anAdvanced Driver Assistance System (ADAS) or support autonomous shuntingoperations. To enable such a system, this paper presents GERALD, a noveldataset for a neural network based detection approach of railway signals.The dataset contains 5000 images from a wide variety of railway scenes aswell as annotations for the most common types of German mainline railwaysignals. The material was gathered using publicly available cab-viewrecordings uploaded by railway enthusiasts on YouTube. Using a state of theart neural network architecture for evaluation, we notice promisingdetection accuracies despite GERALD being a comparably small dataset.

OpenAccess:
ZIP
(additional files)