How do ICP variants perform when used for scan matching terrain point clouds?

Abstract

Many variants of the Iterative Closest Point (ICP) algorithm have been proposed for registering point clouds. This paper explores the performance of 20,736 ICP variants applied to the registration of point clouds for the purpose of terrain mapping, using data obtained from a mobile platform. The methodology of the study has involved taking sequences of 100 consecutive scans at three distinct scenes along the route of a mining haul truck operating in a typical surface mining environment. The scan sequences were obtained at 20 Hz from a Velodyne HDL-64E mounted on the truck. The aim is to understand how well the ICP variants perform in consolidating these scans into sub-maps. Variants are compared against three metrics: accuracy, precision, and relative computational cost. The main finding of the paper is that none of the variants is simultaneously accurate, precise, and fast to compute, across all three scenes. The best performing variants employed strategies that filtered the data sets, used local surface geometry in the form normals, and used the distance between points in one point cloud to a corresponding surface from a reference point cloud as a measure of the fit between two point clouds. The significance of this work is that it: (i) provides guidance in the construction of ICP variants for terrain mapping; and (ii) identifies the significant limitations of existing ICP variants for this application.