Slicing Method for curved façade and window extraction from point clouds

详细信息    查看全文

• 作者：S.M. Iman Zolanvari^a ; ^{Iman.Zolanvari@ucdconnect.ie" class="auth_mail" title="E-mail the corresponding author} ; ; Debra F. Laefer^a ; ^b ; ^{Debra.Laefer@ucd.ie" class="auth_mail" title="E-mail the corresponding author} ;
• 关键词：Laser scanning ; 3D computational modelling ; Window detection ; Feature extraction ; Point cloud data ; LiDAR
• 刊名：ISPRS Journal of Photogrammetry and Remote Sensing
• 出版年：2016
• 出版时间：September 2016
• 年：2016
• 卷：119
• 期：Complete
• 页码：334-346
• 全文大小：7690 K

文摘

Laser scanning technology is a fast and reliable method to survey structures. However, the automatic conversion of such data into solid models for computation remains a major challenge, especially where non-rectilinear features are present. Since, openings and the overall dimensions of the buildings are the most critical elements in computational models for structural analysis, this article introduces the Slicing Method as a new, computationally-efficient method for extracting overall façade and window boundary points for reconstructing a façade into a geometry compatible for computational modelling. After finding a principal plane, the technique slices a façade into limited portions, with each slice representing a unique, imaginary section passing through a building. This is done along a façade’s principal axes to segregate window and door openings from structural portions of the load-bearing masonry walls. The method detects each opening area’s boundaries, as well as the overall boundary of the façade, in part, by using a one-dimensional projection to accelerate processing. Slices were optimised as 14.3 slices per vertical metre of building and 25 slices per horizontal metre of building, irrespective of building configuration or complexity. The proposed procedure was validated by its application to three highly decorative, historic brick buildings. Accuracy in excess of 93% was achieved with no manual intervention on highly complex buildings and nearly 100% on simple ones. Furthermore, computational times were less than 3 sec for data sets up to 2.6 million points, while similar existing approaches required more than 16 hr for such datasets.