Reconstruction of scaffolds from a photogrammetric point cloud of construction sites using a novel 3D local feature descriptor

- An algorithm for detecting scaffolds from the point cloud of the construction site is proposed.
- A framework for reconstructing geometric objects in a construction site is proposed and validated.
- A novel 3D local feature descriptor for delineating linear straight shape objects is designed.

Scaffolds always act as disturbances when reconstructing the 3D scene of the construction site due to occlusions, similarities with buildings in color and height as well as their adjacent positions to wall surfaces. Since scaffolds are commonly utilized to assist the construction and maintenance of building structures, professionals can estimate the overall progress and temporal objects of construction projects by assessing the status or arrangement of the scaffolds. Its thin, repeating and complex structures also make it a valuable dataset for testing related algorithms and approaches for the reconstruction of 3D construction site scene. To this end, we present a data-driven workflow for the detection and reconstruction of scaffolding components, including tubes, toeboards, and decks, given a photogrammetric point cloud. Our workflow consists of two parts: one part concerns the strategy based on projection and methods of grouping and slicing planar surfaces for detecting and extracting points of scaffolds from the construction site. The other part relates to the point feature derivation using a novel 3D local feature descriptor LSSHOT, designed for extracting features in the classification of points. Specifically, our workflow is implemented by five major steps, including preprocessing of the point cloud, division of building facades, classification of points, geometric modeling and refinement of results. To evaluate our proposed descriptor, a series of simulated experiments using synthetic datasets is conducted via shape matching tests. A real application is also carried out to validate the feasibility and effectiveness of our workflow using the photogrammetric point cloud of a construction site. Results of simulated experiments reveal that our proposed descriptor outperforms the original SHOT descriptor in the simulated test, especially when dealing with point clouds having a large percentage of noise. Regarding the real application of reconstructing scaffolds, points of scaffolds are successfully detected, extracted, and reconstructed. For a facade having enough points, over 70% of the scaffolding elements are reconstructed. For the classification of points using LSSHOT descriptor and a random forest classifier, the accuracy of results for the points of two major scaffolding elements reaches more than 70% in our test examples.