Service path localisation from aerial imagery

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Service paths serve as the primary means of accessing railroads and, therefore, must be maintained in optimal condition. However, monitoring all service paths is a daunting task, given their extensive coverage along all railroads in the Netherlands. This research aims to investigate the feasibility of an automatic service path localization system, leveraging recent advancements in computer vision, deep learning, and image segmentation, which have demonstrated impressive results in road segmentation. In order take advantage of these advancements this study selects a segmentation framework, to segment service paths from aerial imagery. In order to be able to train this network a dataset needs to be created as there currently exists no dataset of labelled service paths. Therefore a labelling pipeline is designed. This study also tries to leverage techniques to reduce the required amount of labelled data. The present study employs the U-Net architecture, a successful image segmentation model, as its backbone for segmenting service paths. Transfer learning and augmentation techniques are further employed to alleviate the problem of a limited dataset. The results demonstrate the plausibility of a service path segmentation system, but performance improvements are required before its deployment. To enhance performance, suggested improvements include reducing subjectivity in the labeling pipeline and increasing diversity in the training data.
Faculteit der Sociale Wetenschappen