Relative localization and multi-robot SLAM for autonomous systems

Autonomous systems that navigate or interact with their environment must accurately estimate their own motion and understand their surroundings in order to operate safely and effectively. This cumulative thesis primarily presents scientific contributions in two closely related domains: relative localization and multi-robot simultaneous localization and mapping (SLAM). Relative localization estimates a platform’s motion from arbitrary coordinate frames and forms the foundation of SLAM. SLAM builds on these motion estimates by incorpo- rating exteroceptive sensor data to incrementally construct a model of the environment. SLAM achieves a consistent representation of the surroundings through loop closures, which associate observations from different times and places to reduce accumulated drift. This thesis begins by laying the theoretical and algorithmic foundations of relative lo- calization and SLAM. It introduces key sensors, particularly cameras, light detection and ranging (lidar), and inertial measurement units (IMUs). Mathematical frameworks for state estimation using Kalman filters and graph-based optimization are presented, with a focus on the handling of 3D orientations via the ⊞-operator, which enables state esti- mation on manifolds. Quantitative performance evaluation methods, including absolute and relative trajectory error metrics, are also introduced. The thesis places an emphasis on visual-inertial odometry (VIO) which fuses data from cameras and IMUs for relative localization. In particular, two algorithms designed for the domain of highly automated driving are presented, which additionally use vehicle-related quantities, such as vehicle speed and steering angle information. Inertial state kinemat- ics, which are the backbone of VIO systems, are highlighted and 3D scene projection from camera data is discussed. Further insights into the performance of a contributed VIO algorithm are presented using two IMUs with different performance characteristics. In the domain of multi-robot SLAM, a decentralized graph-based system is proposed that supports various collaborative exploration strategies. The system primarily relies on lidar and integrates robot poses as graph nodes, with spatial constraints encoded as odometry or loop closure edges. The design of the graph structure allows for the flexible integration of sensor data from various sources, which makes it particularly suitable for the cooperation of multiple robotic systems. Further enhancements beyond the orig- inal method are presented, including an offline static map provider, edge removal for erroneous loop closures, and the integration of radio-based distance constraints between mobile robots and stationary beacons. Complementary research efforts are also presented. A high-precision trajectory evalu- ation was performed for an autonomous lawn mower, achieving sub-centimeter accuracy using lidar and an infrared tracking system. This precision was necessary due to the centimeter-level accuracy already provided by global navigation satellite system (GNSS) and real-time kinematic (RTK) localization. Another contribution involves a radar- and lidar-based object tracking framework for highly automated driving applications. Traffic participants, including pedestrians, cyclists, and vehicles, are treated as objects to be relatively localized with respect to the ego vehicle. Special attention is given to the design of an efficient data association method. Overall, this thesis presents both theoretical and practical contributions across multi- ple platforms, including vehicles, autonomous lawn mowers, and ground rovers, advanc- ing the state of the art in optically aided relative localization and SLAM for autonomous systems.