University of Twente Student Theses


Semi-supervised point cloud segmentation on railway data

Dekker, Bram (2023) Semi-supervised point cloud segmentation on railway data.

[img] PDF
Abstract:Monitoring digital twins of the railway infrastructure is safer, less erroneous and faster compared to physical inspection. Digital representations of the railway environment are often obtained via sensors that produce unlabelled point clouds. The point clouds need to be semantically labelled to generate digital twins. Manual labelling requires substantial effort. Deep learning has shown great potential in semantic segmentation tasks using supervised learning. However, there are no publicly available railway datasets. This study thus explores semi-supervised learning for point cloud semantic segmentation. Specifically, two approaches are implemented. In the active learning approach, an algorithm is developed to select the most informative data. The SO-Net segmentation model is trained on only a small portion of the most informative data. For the generative few-shot learning approach, synthetic data is created based on a small labelled dataset. A PointNet++ model is then trained on this synthetic dataset. Both approaches show promising results. The active learning approach achieves over 95% of the performance compared to the fully supervised method using 37.5% less labelled data. The performance of the few-shot learning approach is equivalent to the state-of-the-art while training on only synthetic data. A major challenge for point cloud semantic segmentation in the context of railways is the inherent class imbalance in the data. Further research towards techniques that address class imbalance could improve the performance of the models.
Item Type:Essay (Master)
Saxion University of Applied Sciences, Enschede, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page