Leveraging Incremental Localization Information for Adapting Feature-based Map Matching Methods

Kurzfassung

The current state of the art of SLAM methods are sufficient to enable a specific robot to navigate a specific, static environment. However, the performance strongly depends on an expert to chose the right parameters. This thesis focuses on feature-based map matching, which is often used as a part of a SLAM solution to detect loop closures, thereby reducing the accumulated error of the robot's pose. In practice, the parameter-dependence constrains the robot's autonomy severely, since it can only safely navigate environments of which the rough structure is known beforehand. It also complicates using a shared map in heterogeneous multi-robot teams, because each robot system requires its own set of parameters to reliably navigate its environment. Finally, tuning the many parameters manually is tedious and often quite hard, even with expert knowledge about the system. The high dimensionality of the parameter space and long evaluation durations make tackling this problem with simple approaches, like a grid search, infeasible. To remedy this problem, a framework for automated parameter tuning is proposed that uses Bayesian optimization. It includes a measure for map matcher performance based on both the quality and the quantity of the matches, which only requires ground truth transformations between submap pairs. Labels that tell whether a submap pair should be matchable are not required. The proposed method aims to leverage transformations estimated by a robot's incremental localization system as pseudo ground truth, to calculate the measure during the optimization process. Since most autonomous robots use an incremental localization system, those transformations are available anyway. The proposed solution is evaluated on a simulation dataset and a dataset recorded with a real robot system. The experiments show that the proposed solution is capable of providing map matcher parameters with at least similar performance than a reference parameter set that has been hand tuned for the evaluated datasets.