In construction site progress tracking, a point cloud of the construction site is compared to the 3D BIM-model of the planned building. If a sufficiently large number of points can be detected in the vicinity of a building component, this component is likely built. The schedule, deposited in the BIM-model, is used to derive the overall progress of the construction site and to detect delays. Problems occur when the quality of the point cloud makes the detection of objects ambiguous. Occluded areas, sparse regions, or temporary objects aggravate the comparison between the point cloud and the model. Object detection in point clouds would improve this work ow significantly. Pointwise classification is generally an unsolved problem. The low quality of the point clouds in use, renders the application of classical feature descriptors impossible. Therefore, this thesis investigates how pointwise classification can be applied to point clouds with the help of deep learning. Different deep neuronal network architectures are explored. In order to train the architecture of choice, a dataset is needed. Because there is no shape dataset for construction site related objects, a generator is developed which can generate datasets based on common mesh files. The theory behind machine learning in general and neuronal networks in specific is explained. Followed by the work ow to create a dataset and to train the neuronal network. Results and suggestions for further investigations complete this work.     «

In construction site progress tracking, a point cloud of the construction site is compared to the 3D BIM-model of the planned building. If a sufficiently large number of points can be detected in the vicinity of a building component, this component is likely built. The schedule, deposited in the BIM-model, is used to derive the overall progress of the construction site and to detect delays. Problems occur when the quality of the point cloud makes the detection of objects ambiguous. Occluded area...     »