As the capabilities of autonomous vehicles increase, their use in situations that are dangerous or dull for humans is becoming more popular. Autonomous systems are currently being used in several military and civilian domains, including search and rescue operations, disaster relief coordination, infrastructure inspection and surveillance missions. In order to perform high level mission autonomy tasks, a method is required for the vehicle to localize itself, as well as generate a map of the environment. Algorithms which allow the vehicle to concurrently localize and create a map of its surroundings are known as solutions to the Simultaneous Localization and Mapping (SLAM) problem. Certain high level tasks, such as drivability analysis and obstacle avoidance, benefit from the use of a dense map of the environment, and are typically generated with the use of point cloud data. The point cloud data is incorporated into SLAM algorithms with scan registration techniques, which determine the relative transformation between two sufficiently overlapping point clouds. The Normal Distributions Transform (NDT) algorithm is a promising method for scan registration, however many issues with the NDT approach exist, including a poor convergence basin, discontinuities in the NDT cost function, and unreliable pose estimation in sparse, outdoor environments. This thesis presents methods to overcome the shortcomings of the NDT algorithm, in both 2D and 3D scenarios. To improve the convergence basin of NDT for 2D scan registration, the Multi-Scale k-Means NDT (MSKM-NDT) algorithm is presented, which divides a 2D point cloud using k-means clustering and performs the scan registration optimization over multiple scales of clustering. The k-means clustering approach generates fewer Gaussian distributions when compared to the standard NDT algorithm, allowing for evaluation of the cost function across all Gaussian clusters. Cost evaluation across all the clusters guarantees that the optimization will converge, as it resolves the issue of discontinuities in the cost function found in the standard NDT algorithm. Experiments demonstrate that the MSKM-NDT approach can be used to register partially overlapping scans with large initial transformation error, and that the convergence basin of MSKM-NDT is superior to NDT for the same test data. As k-means clustering does not scale well to 3D, the Segmented Greedy Cluster NDT (SGC-NDT) method is proposed as an alternative approach to improve and guarantee convergence using 3D point clouds that contain points corresponding to the ground of the environment. The SGC-NDT algorithm segments the ground points using a Gaussian Process (GP) regression model and performs clustering of the non ground points using a greedy method. The greedy clustering extracts natural features in the environment and generates Gaussian clusters to be used within the NDT framework for scan registration. Segmentation of the ground plane and generation of the Gaussian distributions using natural features results in fewer Gaussian distributions when compared to the standard NDT algorithm. Similar to MSKM-NDT, the cost function can be evaluated across all the clusters in the scan, resulting in a smooth and continuous cost function that guarantees convergence of the optimization. Experiments demonstrate that the SGC-NDT algorithm results in scan registrations with higher accuracy and better convergence properties than other state-of-the-art methods for both urban and forested environments.

As the capabilities of autonomous vehicles increase, their use in situations that are dangerous or dull for humans is becoming more popular. Autonomous systems are currently being used in several military and civilian domains, including search and rescue operations, disaster relief coordination, infrastructure inspection and surveillance missions. In order to perform high level mission autonomy tasks, a method is required for the vehicle to localize itself, as well as generate a map of the environment. Algorithms which allow the vehicle to concurrently localize and create a map of its surroundings are known as solutions to the Simultaneous Localization and Mapping (SLAM) problem. Certain high level tasks, such as drivability analysis and obstacle avoidance, benefit from the use of a dense map of the environment, and are typically generated with the use of point cloud data. The point cloud data is incorporated into SLAM algorithms with scan registration techniques, which determine the relative transformation between two sufficiently overlapping point clouds. The Normal Distributions Transform (NDT) algorithm is a promising method for scan registration, however many issues with the NDT approach exist, including a poor convergence basin, discontinuities in the NDT cost function, and unreliable pose estimation in sparse, outdoor environments. This thesis presents methods to overcome the shortcomings of the NDT algorithm, in both 2D and 3D scenarios. To improve the convergence basin of NDT for 2D scan registration, the Multi-Scale k-Means NDT (MSKM-NDT) algorithm is presented, which divides a 2D point cloud using k-means clustering and performs the scan registration optimization over multiple scales of clustering. The k-means clustering approach generates fewer Gaussian distributions when compared to the standard NDT algorithm, allowing for evaluation of the cost function across all Gaussian clusters. Cost evaluation across all the clusters guarantees that the optimization will converge, as it resolves the issue of discontinuities in the cost function found in the standard NDT algorithm. Experiments demonstrate that the MSKM-NDT approach can be used to register partially overlapping scans with large initial transformation error, and that the convergence basin of MSKM-NDT is superior to NDT for the same test data. As k-means clustering does not scale well to 3D, the Segmented Greedy Cluster NDT (SGC-NDT) method is proposed as an alternative approach to improve and guarantee convergence using 3D point clouds that contain points corresponding to the ground of the environment. The SGC-NDT algorithm segments the ground points using a Gaussian Process (GP) regression model and performs clustering of the non ground points using a greedy method. The greedy clustering extracts natural features in the environment and generates Gaussian clusters to be used within the NDT framework for scan registration. Segmentation of the ground plane and generation of the Gaussian distributions using natural features results in fewer Gaussian distributions when compared to the standard NDT algorithm. Similar to MSKM-NDT, the cost function can be evaluated across all the clusters in the scan, resulting in a smooth and continuous cost function that guarantees convergence of the optimization. Experiments demonstrate that the SGC-NDT algorithm results in scan registrations with higher accuracy and better convergence properties than other state-of-the-art methods for both urban and forested environments.

This thesis describes and investigates different approaches to indoor mapping and navigation. A system capable of mapping large indoor areas with a stereo camera and/or a laser camera mounted to e.g. a robot or a human is developed. The approaches investigated in this report are either based on Simultaneous Lo- calisation and Mapping (SLAM) techniques, e.g. Extended Kalman Filter-SLAM (EKF-SLAM) and Smoothing and Mapping (SAM), or registration techniques, e.g. Iterated Closest Point (ICP) and Normal Distributions Transform (NDT).In SLAM, it is demonstrated that the laser camera can contribute to the stereo camera by providing accurate distance estimates. By combining these sensors in EKF-SLAM, it is possible to obtain more accurate maps and trajectories compared to if the stereo camera is used alone.It is also demonstrated that by dividing the environment into smaller ones, i.e. submaps, it is possible to build large maps in close to linear time. A new approach to SLAM based on EKF-SLAM and SAM, called Submap Joining Smoothing and Mapping (SJSAM), is implemented to demonstrate this.NDT is also implemented and the objective is to register two point clouds from the laser camera to each other so that the relative motion can be obtained. The NDT implementation is compared to ICP and the results show that NDT performs better at estimating the angular difference between the point clouds.

Robot mapping is a major field of research in robotics. Its basic task is to combine (register) spatial data, usually gained from range devices, to a single data set. This data set is called global map and represents the environment, observed from different locations, usually without knowledge of their positions. Various approaches can be classified into groups based on the type of sensor, e.g. Lasers, Microsoft Kinect, Stereo Image Pair. A major disadvantage of current methods is the fact, that they are derived from hardly scalable 2D approaches that use a small amount of data. However, 3D sensing yields a large amount of data in each 3D scan. Autonomous mobile robots have limited computational power, which makes it harder to run 3D robot mapping algorithms in real-time. To remedy this limitation, the proposed research uses mid-level geometric features (lines and ellipses) to construct 3D geometric primitives (planar patches, cylinders, spheres and cones) from 3D point data. Such 3D primitives can serve as distinct features for faster registration, allowing real-time performance on a mobile robot. This approach works in real-time, e.g. using a Microsoft Kinect to detect planes with 30 frames per second. While previous approaches show insufficient performance, the proposed method operates in real-time. In its core, the algorithm performs a fast model fitting with a model update in constant time (O(1)) for each new data point added to the model using a three stage approach. The first step inspects 1.5D sub spaces, to find lines and ellipses. The next stage uses these lines and ellipses as input by examining their neighborhood structure to form sets of candidates for the 3D geometric primitives. Finally, candidates are fitted to the geometric primitives. The complexity for point processing is O(n); additional time of lower order is needed for working on significantly smaller amount of mid-level objects. The real-time performance suggests this approach as a pre-processing step for 3D real-time higher level tasks in robotics, like tracking or feature based mapping. In this thesis, I will show how these features are derived and used for scene registration. Optimal registration is determined by finding plane-feature correspondence based on mutual similarity and geometric constraints. Our approach determines the plane correspondence in three steps. First step computes the distance between all pairs of planes from the first scan to all pair of planes from the second scan. The distance function captures angular, distance and co-planarity differences. The resulting distances are accumulated in a distance matrix. The next step uses the distance matrix to compute the correlation matrix between planes from the first and second scan. Finally plane correspondence is found by finding the global optimal assignment from the correlation matrix. After finding the plane correspondence, an optimal pose registration is computed. In addition to that, I will provide a comparison to existing state-of-the-art algorithms. This work is part of an industry collaboration effort sponsored by the National Institute of Standards and Technology (NIST), aiming at performance evaluation and modeling of autonomous navigation in unstructured and dynamic environments. Additional field work, in the form of evaluation of real robotic systems in a robot test arena was performed. / Computer and Information Science / Accompanied by two .mp4 files.

A large number of current mobile robots use 3D sensors as part of their sensor setup. Common 3D sensors, i.e., laser scanners or RGB-D cameras, emit a signal (laser light or infrared light for instance), and its reflection is recorded in order to estimate depth to a surface. The resulting set of measurement points is commonly referred to as 'point clouds'. In the first part of this dissertation an inherent problem of sensors that emit some light signal is addressed, namely that these signals can be reflected and/or refracted by transparent of highly specular surfaces, causing erroneous or missing measurements. A novel heuristic approach is introduced how such objects may nevertheless be identified and their size and shape reconstructed by fusing information from several viewpoints of the scene. In contrast to other existing approaches no prior knowledge about the objects is required nor is the shape of the reconstructed objects restricted to a limited set of geometric primitives. The thesis proceeds to illustrate problems caused by sensor noise and registration errors and introduces mechanisms to address these problems. Finally a quantitative comparison between equivalent directly measured objects, the reconstructions and "ground truth" is provided. The second part of the thesis addresses the problem of automatically determining the quality of the registration for a pair of point clouds. Although a different topic, these two problems are closely related, if modeled in the fashion of this thesis. After illustrating why the output parameters of a popular registration algorithm (ICP) are not suitable to deduce registration quality, several heuristic measures are developed that provide better insight. Experiments performed on different datasets were performed to showcase the applicability of the proposed measures in different scenarios.