Multiple 3D scans are necessary to digitalize environments without occlusions. To create a correct and consistent model, the scans have to be merged into one coordinate system. This process is called registration. If the localization of the robot with the 3D scanner were precise, the registration could be done directly by the robot pose. However, due to the unprecise robot sensors, the self localization is erroneous, so the geometric structure of overlapping 3D scans has to be considered for registration.
The matching of 3D scans can either operate on the whole 3D scan point set or can be reduced to the problem of scan matching in 2D by extracting, e.g., a horizontal plane of fixed height from both scans, merging these 2D scans and applying the resulting translation and rotation matrix to all points of the corresponding 3D scan.
Matching of complete 3D scans has the advantage of having a larger set of attributes (either pure data points or extracted features) to compare the scans. This results in higher precision and lowers the possibility of running into a local minimum of the cost function. Furthermore, using three dimensions enables the robot control software to recognize and take into account changes of height and roll, yaw and pitch angles of the robot. This 6D robot relocalization is essential for robots driving cross country or in mines.
6D matching approaches of 3D surfaces can be classified into two categories: First, scan matching as optimization problem uses a cost function for the quality of the alignment of the scans. The range images are registered by determining the rigid transformation (rotation and translation) which minimizes the cost function. Second, feature based scan matching extracts distinguishing features of the range images and uses corresponding features for calculating the alignment the scans. Even though through this approach is more intuitive, it cannot be applied to scan matching in mines, since the surface structure of the mine is too simple. In consequence there are not many features and an algorithm based on feature matching will fail [22,17].