Robust State Estimation and Mapping for Autonomous Inspection

Mobile robots are increasingly used for autonomous exploration and inspection because of the advancements in the technology of robot autonomy. In this thesis, we focus on the state estimation and mapping problem in the context of robot autonomous inspection. Although there exist well-established foundations, applying the technology in the real world remains challenging. Therefore, we are particularly interested in the robustness of algorithms when being applied in the real world. We believe addressing the challenges that arise from inspection-specific tasks will have more general impacts on other applications. Specifically, our work contains the following 4 contributions: Firstly, we present a LiDAR-based pose tracking algorithm to achieve real-time robust state estimation. The proposed method is formulated using a probabilistic filtering framework named Error State Kalman Filter (ESKF) and can be used on lowend computing devices to achieve a robust estimation of agile motions. Despite the successful deployment in various environments, the method finds difficulties in scenarios with limited geometric features, such as long straight tunnels. Secondly, we think about the failure cases of general LiDAR-based state estimation algorithms in situations where the geometric features are limited. A localizability model is presented and can be used to predict the confidence of localization based on the analysis of sensor readings. We then demonstrate successful localization in a real tunnel by fusing sensors with complimentary localizability. Thirdly, we turn our attention to dense reconstruction for inspection. Inspired by the idea of complementary sensing, we combine LiDAR and camera data in a joint optimization framework to build globally consistent and locally dense 3D maps. Additionally, the proposed joint optimization framework is used to solve large-scale realworld problems by enhancing Structure from Motion (SfM) algorithms with LiDAR data providing structural constraints. Using the proposed methods, we demonstrate successful dense reconstruction of a real bridge tunnel. Lastly, having noticed the structured nature of reconstructed dense models, we explore the possibility of using geometric primitives to build high-level compact maps. More specifically, we present a quadrics-based representation to unify multiple types of primitives in man-made environments, which leads to a concise and geometrically meaningful SLAM formulation. We show that using geometric primitives is not only a compact choice, the encoded high-level information also benefits the motion estimation by providing stronger constraints.