Automated segmentation of in-service infrastructure datasets acquired using laser scanning data

  1. Lamas Novoa, Daniel
Dirixida por:
  1. Belén Riveiro Rodríguez Director
  2. Mario Soilán Rodríguez Director

Universidade de defensa: Universidade de Vigo

Fecha de defensa: 20 de marzo de 2024

Tribunal:
  1. Lucía Díaz Vilariño Presidenta
  2. Sander Jakob Oude Elberink Secretario/a
  3. Luigi Barazzetti Vogal

Tipo: Tese

Resumo

The relevance of transportation infrastructure in our societies, as part of Critical Infrastructure Systems (CIS), indicates the importance of monitoring it. To monitor these assets, as-is models can be created. Among the most important transport infrastructures, some are linear infrastructures with specific characteristics that must be considered for their modelling. To accurately represent them, it is necessary to gather information about their current state without disrupting their function. LiDARs are a commonly used technology for obtaining information of linear transport infrastructures. They provide geometrical and geo-referenced information in form of point clouds. In this context, this doctoral dissertation proposes different methodologies to contribute to the field of knowledge of methods that allow the generation of geometric models of linear transport infrastructure from geo-referenced data that provide information about the as-is asset. To achieve this, we investigate how point clouds can be automatically segmented (semantically and in instances) to extract relevant information for modelling them. Various artificial intelligence segmentation techniques are explored, and segmentation algorithms are developed based on them, while studying their advantages and disadvantages. The most relevant characteristic of these algorithms to focus on is their ability of generalise or be readapted to different scenarios. The algorithms developed in this doctoral dissertation can be divided into heuristic and deep learning algorithms. Deep learning methods often require large amounts of data for training, which may not always be readily available. Therefore, we have also explored the creation of synthetic data and studied its impact on the training process. The algorithms presented in this doctoral dissertation have been tested in various real-world scenarios, such as railway, road, and truss-type bridge environments. This resulted in three publications in high-impact, peer-reviewed, international journals indexed on the Journal Citation Report (JCR) and one conference paper. These publications advanced the state of the art and contributed to the knowledge in their respective field.