The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLI-B3
https://doi.org/10.5194/isprs-archives-XLI-B3-725-2016
https://doi.org/10.5194/isprs-archives-XLI-B3-725-2016
10 Jun 2016
 | 10 Jun 2016

CLASSIFICATION OF PHOTOGRAMMETRIC POINT CLOUDS OF SCAFFOLDS FOR CONSTRUCTION SITE MONITORING USING SUBSPACE CLUSTERING AND PCA

Y. Xu, S. Tuttas, L. Heogner, and U. Stilla

Keywords: Scaffolding components, Photogrammetric point clouds, Subspace clustering, PCA, Classification

Abstract. This paper presents an approach for the classification of photogrammetric point clouds of scaffolding components in a construction site, aiming at making a preparation for the automatic monitoring of construction site by reconstructing an as-built Building Information Model (as-built BIM). The points belonging to tubes and toeboards of scaffolds will be distinguished via subspace clustering process and principal components analysis (PCA) algorithm. The overall workflow includes four essential processing steps. Initially, the spherical support region of each point is selected. In the second step, the normalized cut algorithm based on spectral clustering theory is introduced for the subspace clustering, so as to select suitable subspace clusters of points and avoid outliers. Then, in the third step, the feature of each point is calculated by measuring distances between points and the plane of local reference frame defined by PCA in cluster. Finally, the types of points are distinguished and labelled through a supervised classification method, with random forest algorithm used. The effectiveness and applicability of the proposed steps are investigated in both simulated test data and real scenario. The results obtained by the two experiments reveal that the proposed approaches are qualified to the classification of points belonging to linear shape objects having different shapes of sections. For the tests using synthetic point cloud, the classification accuracy can reach 80%, with the condition contaminated by noise and outliers. For the application in real scenario, our method can also achieve a classification accuracy of better than 63%, without using any information about the normal vector of local surface.