Automated reconstruction of Building Information Model objects from point cloud data

As-built Building Information Models are becoming increasingly popular in the Architectural, Engineering and Construction industry. These models reflect the state of a building up to as-built conditions and are used in numerous applications such as refurbishment, facility management and project planning. The production of these models includes the acquisition of the structure with remote sensing techniques and the manual placement of the target objects. However, the interpretation of the acquired point cloud data is time-consuming and prone to misinterpretations. Aside from the overwhelming data size, modelers struggle with occlusions, noise and clutter in the building environment. This work presents a series of automated procedures to cope with the current problems. More specifically, both theoretical and practical solutions are proposed to aid in the data acquisition, point cloud interpretation and object reconstruction. The following contributions are part of this research. Data Acquisition The purpose of the data acquisition is to produce highly dense accurate point cloud data. An extensive literature study is performed to explore the opportunities of both static and dynamic data acquisition systems for building surveying. It is concluded that while mobile solutions have superior scene coverage and acquisition speed, terrestrial laser scanners currently have the upper hand in producing high-quality point clouds fit for as-built modeling. A practical study proves that the current workflow, which includes control measurements with total stations, can be significantly enhanced by integrating compensator data in the registration of the individual scans. One of the major conclusions is that terrestrial laser scanning can be used as a standalone solution in mid-to-large scale projects without the need for control measurements, while still producing data conform the metric accuracy requirements. Data Interpretation Once the data is aligned in a common coordinate system, the point clouds are interpreted to determine which points belong to the objects of interest. The goal is to make a proper interpretation of the data to increase the level of information. Instead of a manual procedure, a fully automated workflow is proposed which consists of three consecutive steps. First, the data is segmented into planar primitives according to the surface hypothesis of the structural elements in the observed data. Next, each of the observed segments is assigned a predefined semantic class label through a pretrained machine learning algorithm. The resulting classified instances are clustered together using graph theory in order to isolate all the available observations of each relevant object. From the experiments it could be derived that the majority of the observations of structural objects in a building environment can be properly extracted. While some false positives remain a problem, the implementation of a fully 3D approach shows promising results for point cloud interpretation. Data Reconstruction The grouped observations of each object are used to reconstruct a set of generic Building Information Modeling entities conform existing standards. The target is to create a set of parametric BIM objects that are usable by the industry. A class specific reconstruction algorithm is proposed to extract the necessary properties for the parametric representation of the walls. After the creation of a set of partial objects, the topology is also reconstructed based on intersection theory. Overall, the experiments show that despite the abstractions of the class definitions, the automated workflow is capable of reconstructing topologically consistent wall geometry.status: publishe