Registration between Multiple Laser Scanner Data Sets

Abstract

Laser scanners provide a three-dimensional sampled representation of the surfaces of objects with generally a very large number of points. The spatial resolution of the data is much higher than that of conventional surveying methods. Since laser scanners have a limited field of view, it is necessary to collect data from several locations in order to obtain a complete representation of an object. These data must be transformed into a common coordinate system. This procedure is called the registration of point clouds. In terms of input data, registration methods can be classified into two categories: one is the registration of two point clouds from different scanner locations, so-called pair-wise registration (Rusinkiewicz & Levoy, 2001), and the other is simultaneous registration of multiple point clouds (Pulli, 1999; Williams & Bennamoun, 2001). However, the global registration of multiple scans is more difficult because of the large nonlinear search space and the huge number of point clouds involved. Commercial software typically uses separately scanned markers that can be automatically identified as corresponding points. Akca (2003) uses the special targets attached onto the object(s) as landmarks and their 3-D coordinates are measured with a theodolite in a ground coordinate system before the scanning process. Radiometric and geometric information (shape, size, and planarity) are used to automatically find these targets in point clouds by using cross-correlation, the dimension test and the planarity test. According to the automatic registration problems, several efforts have been made to avoid the use of artificial markers. One of the most popular methods is the Iterative Closest Point (ICP) algorithm developed by Besl & McKay(1992) and Chen & Medioni (1992). ICP operates two point clouds and an estimate of the aligning rigid body transform. It then iteratively refines the transform by alternating the steps of choosing corresponding points across the point clouds, and finding the best rotation and translation that minimizes an error metric based on the distance between the corresponding points. One key to this method is to have a good priori alignment. That means, for partially unorganized and overlapping points, if there is lack of good initial alignment, many other ICP variant don’t work well because it becomes very hard to find corresponding points between the point clouds. Also, although a considerable amount of work on registration of point clouds from laser scanners, it is difficult to understand convergence behavior related to different starting conditions, and error metrics. Many experiment showed that the rate of convergence of ICP heavily relies on the choice of the corresponding point-pairs, and the distance function.