Evaluation of registration methods for sparse 3D laser scans

Abstract

The registration of 3D laser scans is an important task in mapping applications. For the task of mapping with autonomous micro aerial vehicles (MAVs), we have developed a light-weight 3D laser scanner. Since the laser scanner is rotated quickly for fast omnidirectional obstacle perception, the acquired point clouds are particularly sparse and registration becomes challenging. In this paper, we present a thorough experimental evaluation of registration algorithms in order to determine the applicability of both the scanner and the registration algorithms. Using the estimated poses of the MAV, we aim at building local egocentric maps for both collision avoidance and 3D mapping. We use multiple metrics for assessing the quality of the different pose estimates and the quality of the resulting maps. In addition, we determine for all algorithms optimal sets of parameters for the challenging data. We make the recorded datasets publicly available and present results showing both the best suitable registration algorithm and the best parameter sets as well as the quality of the estimated poses and maps.