Towards detection of cracks and their condition in Dutch railway sleepers with Computer Vision
Keywords
Loading...
Authors
Issue Date
2022-08-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
ProRail is the Dutch railway infrastructure manager. To monitor the condition of the
railroad, it currently has to be inspected manually. This is an enormous amount of
work, not always safe and may disrupt train traffic. Besides, to account for increasing
transport and speeds over the coming years, the railway will need to be inspected for
degeneration even more regularly. Therefore, ProRail aims to use a data-driven approach
and detect anomalies in the railroad automatically. To do so, the mobile video inspection
units, their camera-equipped trains, make images of the Dutch railroad. In this research
project we investigate how to use computer vision on the images of the mobile video
inspection units of ProRail to create segmentations of cracks in sleepers and to see what
information can be extracted from the found cracks. A model was developed featuring
a masked R-CNN architecture to perform crack segmentation. To pre-train this model
two external datasets were used, containing 118 and 445 images and their binary masks.
Subsequently, the model pre-trained on external data, was trained on an internal dataset
of 92 images of cracked sleepers from the Dutch railway. The performance of the best
scoring model was an IoU of 0.310 and an AUC of 0.880. Thereby, it was not able to
match the AUC of 0.945 accomplished by a similar study by Li et al. [11]. However,
there are a few reasons that can explain this gap. First of all, the data supplied by
the mobile video inspection units has poorer lighting and resolution than the data used
by Li et al. Secondly, not a lot of images with cracked sleepers are identified. This
made training the masked R-CNN more challenging. Lastly, the AUC was calculated
on a validation set containing more edge cases than usual. Still, even with only a small
internal dataset of 94 samples of cracked sleepers, this research project shows to be
close to a well-performing crack segmentation system that can be applied to the Dutch
railroad. Yet, before such a system can be deployed, first it is paramount to increase the
size of the internal dataset with cracked sleepers and improve the lighting and resolution.
This would result in better performance, making it possible for ProRail to keep track of
the length and thickness of cracks, thereby monitoring sleeper decay over time.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen