Scalable and Reliable Classification of Tunnels using Advanced Machine Learning Algorithms on Point Clouds

ABSTARCT Building Information Modeling (BIM) has emerged as a powerful tool for managing and maintaining infrastructure, providing rich contextual information that can facilitate more effective decision-making. However, when it comes to tunnels, the absence of detailed BIM models is a common challenge, primarily due to the lack of available information. A major challenge in developing comprehensive BIM models for existing tunnels is the need for accurate and reliable survey data. In the past, surveying techniques were used that led to days and days of surveying and led to high costs. Today, however, Terrestrial Laser Scanning (TLS) is adopted instead, leading to a more precise result and shorter data acquisition times. The aim of this thesis is to optimize and enhance the processes starting from data acquisition leading to the classified point cloud. In order to achieve this objective, a novel algorithm has been developed. This algorithm builds upon existing methods for the classification and segmentation of tunnel point clouds, incorporating advanced machine learning techniques such as Ransac, Dbscan, Decision Tree and Random Forest. By leveraging rich contextual information from BIM databases during the classification process, this algorithm ensures high levels of accuracy while maintaining scalability for large datasets. A critical challenge in developing such an algorithm is the need to effectively balance the competing demands of accuracy and scalability. On the one hand, it is essential that the algorithm is able to accurately classify galleries in order to provide meaningful insights into their characteristics and behavior. On the other hand, it is equally important that the algorithm is able to scale effectively to large datasets, in order to ensure that it can be applied in with all types of tunnel-related point cloud acquisitions, not limited to existing cloud or cloud segment subsets within the training model. In order to attain this balance, a combination of supervised and unsupervised learning methods has been utilized. Supervised learning methods are used to train the algorithm on labeled data, allowing it to accurately predict the classification of galleries based on their observed characteristics. Subsequently, unsupervised learning techniques are employed to extract significant characteristics from the gallery data, enabling the algorithm to recognize patterns and connections that may not be readily discernible from the unprocessed data. Overall, this algorithm represents a significant step forward in the development of reliable and scalable gallery classification methods. By leveraging advanced machine learning techniques and incorporating rich contextual information from BIM databases, this approach is able to achieve high levels of accuracy while remaining scalable to large datasets. It is believed that this work has the potential to make a significant contribution to the field of tunnel maintenance and management, providing valuable insights into the behavior of existing tunnels and facilitating more effective decision-making.