3-D Scan Registration Using Normal Distributions Transform with Supervoxel Segmentation

Abstract

In order to use mobile robots in various applications, such as exploration, rescue, surveillance, and military, the most fundamental required capability is an autonomous navigation.Moreover, one of the most important problem of autonomous navigation is that a robot should build a map of its surroundings and identify its location on its own map, i.e., a simultaneous localization and mapping (SLAM) problem.The information about the surroundings of a robot, which is used in SLAM algorithms, has two types, sparse features and dense point clouds.However, in order to perform a detailed path planning and collision avoidance, a map with dense point clouds is necessary because dense point clouds have rich information on the surrounding obstacles.Also, the map has to be three-dimensional (3-D) so that various shapes of robots carry out wide-ranging tasks.Therefore, the SLAM algorithms using dense point clouds are required for an autonomous navigation, but in order to guarantee the performance of SLAM, a high performance 3-D scan registration algorithm is essential.This thesis presents what is termed the supervoxel normal distributions transform (SV-NDT), a novel three-dimensional registration algorithm which improves the performance of the three-dimensional normal distributions transform (3-D NDT) significantly.The 3-D NDT partitions a model scan using a 3-D regular grid.However, generating normal distributions using the 3-D regular grid causes considerable information loss because the 3-D regular grid does not use any information pertainii ing to the local surface structures of the model scan.The best type of surface (the constituent unit of each scan) for modeling with one normal distribution is known to be the plane.The SV-NDT reduces the loss of information using a supervoxelgenerating algorithm at the partitioning stage.In addition, it uses the information of the local surface structures from the data scan by replacing the Euclidean distance with a function that uses local geometries as well as the Euclidean distance when each point in the data scan is matched to the corresponding normal distribution.Experiments demonstrate that the use of the supervoxel-generating algorithm increases the modeling accuracy of the normal distributions and that the proposed 3-D registration algorithm outperforms the 3-D NDT and other widely used 3-D registration algorithms in terms of robustness and speed on both synthetic and realworld datasets.Additionally, the positive effect of changing the function to create correspondences on the performance of registration is also verified.