Planar Segments Based Three-dimensional Robotic Mapping in Outdoor Environments

Abstract

This dissertation focuses on the problem of three-dimensional (3D) outdoor robotic mapping. Laser scanners are chosen as the primary sensors and a novel approach for scan registration based on planar segments is developed. Unlike most existing approaches, it does not require an a-priori pose estimation from other sensors such as odometers and inertial measurement units (IMUs). Instead, the transformation is determined globally by searching corresponding planar segments between overlapping scans.

There are three steps in the approach: The first step is to segment each point cloud into planar segments. Depending on the clutter level of the environment and whether the point cloud is organized, four complementary strategies have been proposed, namely a point based, a subwindow based, a hybrid and a cached octree region growing algorithm. Among them, the former three are limited to organized point clouds; in addition, the second one is restricted to structured environments. Based on observations from field experiments, the hybrid/cached octree region growing algorithms are recommended for organized/unorganized point clouds.

The second step is to calculate the area of each segment resulting from the first step. Again, segments from organized and unorganized point clouds are distinguished, where an alpha-shape based algorithm is proposed for unorganized point sets and a range-image based method is proposed for organized point sets, respectively.

The third step is to find segment correspondences and compute the transformation based on matched segments. The correspondences are searched globally in order to maximize a spherical-correlation-like metric, wherein the search space is pruned by both self-similarity and interrelations (geometric constraints). The novelty of the search algorithm is that only the area and plane parameters of each segment are required.

Four datasets acquired by scanners with different field of views have been used to evaluate the proposed approach: three are publicly available and one stems from our custom-built platform. Based on these datasets, the following evaluations have been done: segmentation speed-benchmarking, segment area calculation accuracy- and speed-benchmarking, and registration accuracy comparison with ground-truth. Also, the robustness of the approach with respect to occlusions and partial observations has been proven. The approach has been compared to the Iterative Closest Point (ICP) and Minimum Uncertainty Maximum Consensus (MUMC) algorithms; furthermore, it has been successfully extended to the domain of map merging. Experimental results confirm that the approach offers an alternative to state of the art algorithms in plane-rich environments.