Localization And Mapping For Mobile Robot In Large-scale Unstructured Environment

For most of the robotics applications, it is sophisticated to obtain the pre-knowledge of the environment which is significant for path planning, motion pre-diction and robot control. Therefore, Simultaneous Localization and Mapping(SLAM), estimating the pose of robot and mapping the environment concur-rently, plays an important role in mobile robot research. The SLAM systems areimplemented in a number of fields successfully, such as, planet exploration, un-derwater reconnaissance, mining automation, unmanned aerial vehicle navigationand disaster rescue.Most of the previous works focused on the small-scale structured environ-ment. It is widely accepted by the SLAM community that developing the reliableexperimental platform and algorithm system for large-scale unstructured case arethe challeges. This thesis focuses on the SLAM solution to car-liked robot rovingin urban area. it develops 2D laser-only hardware as well as software systems andtackles the state estimate issues with only laser point clouds. Furthermore, toimprove the tractability, the odometry, widely employed in conventional works,are not adopted. Instead, these functions are realized by corresponding algorithmcomponents. The main works of the thesis are addressed as follows:Firstly, the fundamentals and key issues of SLAM are investigated. Thetheory framework, probabilistic basis and prospects are addressed. The stateformulation, map description and environment modelling strategies are studied.And the advantages of information filtering over Kalman filtering in the contextof SLAM are discussed.Secondly, the approaches to the estimation of robot's relative motion withlaser range finder are studied. The problem definition and solutions to laserscan alignment are addressed. Moreover, the classical scan matching algorithm, Iterative Closest Point (ICP), is summarized and demonstrated with real data.As a result, ICP's drawbacks are presented and the potential improvements arediscussed.Thirdly, in order to solve the practical SLAM problem in large-scale unstruc-tured environment, the laser point association is inferred with machine learningapproach. The scan matching algorithm based on Conditional Random Fields(CRF) graph is explored and implemented. The graphical model building, theparameter learning and the probability inference for the 2D laser point cloudsare tackled. To achieve strong robustness and strict probabilistic basis, the multigeometric feature extractions and intelligent management of them are proposed.All these points make the uncertainty quantification and the construction of fil-tering system possible.Fourthly, the uncertainty estimation algorithm for laser scan matching basedon CRF is proposed. The uncertainty sources are investigated and revealed.Particularly, the mechanism of the main uncertainty source is discussed. Thenthe insight of uncertainty tracking into the procedure of probabilistic inferenceis made. The cumulative probability sampling (CPS) is performed in messageconstruction of the belief propagation (BP). And the uncertainty distributionis mapped from the laser point association space to relative motion estimatespace. Eventually, the Sampling-product uncertainty inference algorithm is pro-posed. Compared with traditional algorithms, the Sampling-product inferencedeals with the main uncertainty source more e?ciently. And it utilizes the proba-bilistic proofs stemmed from the inference procedure based on geometric featuresof the laser points. This employs a strict probability foundation. Both the sim-ulation and real robot experiments demonstrate that the uncertainty quantifiedby the proposed algorithm is more reasonable. It is able to work with the fil-tering system to correct the process modelling error. Moreover, the proposedSampling-product uncertainty inference algorithm can be extended to other dataassociation applications other than laser scan matching.Fifthly, the experimental platform and algorithm system are proposed forlarge-scale SLAM in unstructured urban environment. The field tests are carriedout at the University of Sydney (USYD) close to Australian Center for FieldRobotics (ACFR). The trajectory of the robot is longer than 1km. In the sys- tem, the exactly sparse delayed-state information filtering is implemented. Andit combines the CRF-based scan matching, view-based SLAM framework andSampling-product uncertainty inference. As a crucial component, the loop clo-sure strategy for the unstructured outdoor cases is proposed. The field testsdemonstrate that the proposed SLAM system works well with the practical en-vironment and outperforms over the conventional algorithm.The SLAM system developed in this thesis facilitates the autonomous mobilerobot application in the complicated circumstance. It is of great importance toboth military and civil application.