From point cloud to CAD model
Problem Statement:
The Netherlands has about 3000km of railway track, every 50-70m there will be a catenary arch which carry the power lines above the track. There are many variations of catenary arches used in the Netherlands, they consist of a mix of custom components and legacy components. If repairs or renovation is required of these arches, first an assessment is made manually. Goal of this project is to automate this process. To do so, high resolution point cloud scans of catenary arches have been collected using a mobile laser scanning device.
The current method depends on labelled data to train a machine learning model which is capable of segmenting and classifying the point cloud scans of catenary arches. This labelling process is done manually and is tedious and time consuming. This assignment will look into an alternative approach to perform this task and will circumvent the labelling issue. This alternative approach leverages the fact that a library of CAD components exists.
Tasks:
The envisioned approach is as follows:
- CAD library components to point clouds
- Derive representative features using deep learning
- Match CAD library features to collected cloud point features
- Fine-tune match using RANSAC/ICP to get an exact component position within the input point cloud
This approach will circumvent the labelling tasks and will enable the quick addition of new components. An issue which still remains is the large size variation within the scene, an insulator measures 80cm whilst a top bar measures 24m. This is a factor 80 difference in size.
Work:
15% Theory, 65% Practical, 20% Writing
Personal thoughts and ideas are encouraged. The assignment is not set in stone!
Contact:
Bram Ton <b.t.ton@saxion.nl>