University of Twente Student Theses
Automatic recognition of underground utilities in point clouds
Kruiper, J.A. (2024) Automatic recognition of underground utilities in point clouds.
|
PDF
3MB |
| Abstract: | Underground utilities are crucial for economic development. Stakeholders involved in the installation of new utilities include network users, utility companies, landowners, and contractors. Contractors like Siers Groep provide services in underground infrastructure, including accurate mapping and registration of cables and pipelines. The process of mapping underground utilities can be more efficient. One of the novel methods to map utilities is to make a 3D point cloud scan of a trench. Processing the point clouds is still laborious because it is a manual task. Workers at surveying companies do not necessarily have the experience with such a task but also would have a significant amount of additional work in case they need to process 3D point clouds. Automating the process of retrieving the location data of utilities in point clouds enhances efficiency significantly. Automation can be done using machine learning. The objective of this research is to compare and use machine learning algorithms for automatically recognizing and retrieving underground utilities that are present in 3d point clouds of open trenches and convert the data into geometric shapes. From the objective, one main research question is formulated: • How can different machine learning algorithms be used to retrieve utilities from point clouds of open trenches? The main research question is answered by four sub-questions. These are: 1. What types of machine learning algorithms are most useful? 2. How can the machine learning algorithms be used? 3. How should the point cloud data be pre-processed? 4. How can the quality of a model be assessed? To answer the four sub-questions, first different machine learning algorithms are identified and assessed on a set of criteria. After that, the point cloud data used for this research is pre-processed. In this step, a train and test dataset is created including point clouds representing ‘utility’ and ‘not utility’. From the assessment of the identified machine learning algorithms, three algorithms are chosen. The chosen algorithms are trained on the point cloud training data. The result is a model. To choose the best model for the problem, the quality of the models is assessed. The best model is used for a demonstration case, in which a new point cloud is classified. Different supervised machine learning algorithms are identified as potential candidates. Three algorithms are chosen to be trained. These are VoxNet, a shallow neural network and PointNet. These algorithms score best on the set of criteria. Random forests, support vector machines, PointNet ++ and the combination of UnitNet, FeatureNet and FinalNet are not trained. The algorithms can be used via different programs. For this thesis, MATLAB is used because there is clear documentation available about using the algorithms in MATLAB. All three models are trained based on a set of training options. The three models are trained on 10 different combinations of training options. For each model, one option with the best training options is chosen. The three models are then compared and the best model is chosen. This is VoxNet. This model is used for a demonstration case. In the demonstration case, a new point cloud is classified. The utilities classified in this point cloud are converted to polylines that can be used for mapping. In conclusion, the study contributes to the existing literature on machine learning algorithms' application in underground utility recognition from point cloud data. By comparing and evaluating various algorithms, it provides insights into their suitability for utility recognition tasks. Overall, the study highlights the effectiveness of machine learning algorithms in utility recognition from point cloud data. |
| Item Type: | Essay (Bachelor) |
| Faculty: | ET: Engineering Technology |
| Programme: | Civil Engineering BSc (56952) |
| Link to this item: | https://purl.utwente.nl/essays/98650 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page