In the field of Scan-to-BIM, recent developments achieve promising results in accuracy and flexibility, leveraging tools from the field of deep learning for semantic segmentation of raw point cloud data. Those methods demand large-scale, domain-specific datasets for training. Promising ideas to fulfill this need use primitive synthetic point cloud data, which predominantly lack distinct point cloud properties, such as missing patches due to occlusions in the scene. To solve this issue, we use a specialized laser scan simulation tool from the domain of Geosciences in a toolchain that allows generating realistic ground truth data based on 3D models. In this context, we introduce a comprehensive taxonomy for the industrial point cloud context. Furthermore, we provide the missing link for a comprehensive, open-source toolchain that is flexible towards any use case in the field.

Even when adherence to project schedule is the most critical performance metric among project owners, still 53 % of typical construction projects exhibit schedule delays. To contribute to more efficient construction progress monitoring, this research proposes a method to detect the most common temporary object classes in large-scale laser scanner point clouds of construction sites. The proposed workflow includes a combination of several techniques: image processing over vertical projections, finding patterns in 3D detected contours, and performing checks over vertical cross-sections. A deep learning algorithm was leveraged to classify these cross-sections for the purpose of formwork detection. After applying the method on three real-world point clouds and testing with three object categories (cranes, scaffolds, and formwork), the results reveal that the process achieves average rates above 88 % for precision and recall and outstanding computational performance (1 s to process 10 5 points). These metrics demonstrate the method's capability to support the automatic segmentation of point clouds of construction sites.