An Automatic Robust Point Cloud Registration on Construction Sites (original) (raw)

2017, International Workshop on Computing for Civil Engineering (IWCCE)

Point cloud-based reconstruction is a crucial process for process monitoring and as-built modeling, but still a challenging task in construction applications. Due to the finite perspective of a scan, multiple scans from different locations are required to obtain full scene coverage. Then, a registration process is required to transform all point clouds into a common coordinate frame. In this paper, a fully automated target-less point cloud registration method is presented. For this approach, a laser scanning device and a co-registered digital camera are required. This method consists of three major steps for an automatic registration of point clouds. First, it detects common features on the corresponding overlapped images from different locations. Second, it achieves coarse registration using the extracted features on images. Lastly, the fine registration is attained by the iterative closet point (ICP) algorithm. The proposed novel method was tested on a real construction site and the test results of automatic registration of all different scan positions were successfully verified in terms of accuracy. INTRODUCTION Both laser scanning and photogrammetry have been highly studied with the recent improvement in sensing technologies. These two technologies have strengths and weaknesses based on working environments and data quality requirements; however, photogrammetry cannot provide the same level of accuracy as laser scanners do in general. In particular, 3D laser scanning technology has been widely used in construction applications to render as-built objects or environments in the form of dense point cloud data. The abundant amount of point cloud data can be used to efficiently model a construction site. A complete point cloud data of the construction site without shading area can be acquired by scanning multiple times in different scanning positions. Each scan position defines a local coordinate system, which means that the point cloud of each viewpoint is referred to this local frame. A common reference system is required to analyze different laser scanning positions. The transformation process of all local coordinate systems into a common reference coordinate system is called registration. Many studies have been conducted on the registration of multi-view point clouds in the past few years. There are several types of point cloud registration have been studied; target-based, ICP-based, and feature-based. Target-based registration is a reliable and precise method, but is time-consuming. Before scanning, the targets must be placed inside the scanning area. After scanning, the targets have to be collected and manually or automatically recognized to define corresponding points for the registration. Becerik-gerber et al. (2011) proposed a target-based point cloud registration method. They experimented with three different types of targets such as fixed paper, paddle, and sphere, and with