Application of a Performance Assessment Method for Rail Vehicle Localization Algorithms

Abstract

On-board localization systems are becoming increasingly important for modern rail transport. To develop the map-supported sensor data fusion algorithms used for this purpose, expensive ground truth reference systems are currently required to verify the robustness and reliability of these algorithms. This evaluation approach is costly and time-consuming. This thesis therefore presents a method for the systematic analysis of these on-board localization algorithms. A software-based modular framework is introduced that applies artificial degradation to real sensor data without relying on external reference data. This enables the investigating of how algorithms handle various sources of error and what impact this has on the algorithm performance. To demonstrate the applicability of the method, a map-supported Kalman filter bank serves as a concrete use case. The results reveal that the filter bank is capable of partially compensating for stochastic noise through a targeted adjustment of its filter parameters. However, in the presence of systematic errors, the filter bank’s adjustment fails, resulting in significant position deviations while causing the algorithm to overestimate its own accuracy. In summary, this thesis demonstrates that the developed testing method serves as a highly scalable and resource-efficient tool for future algorithm development and testing, as it reduces the need for costly ground truth data. Furthermore, the results illustrate that continuous online calibrations and proactive detection mechanisms will be crucial for reliable localization in rail transport in the future. With the help of such upstream measures, systematic errors can be detected early, preventing a dangerous overestimation of the localization accuracy while supporting a reliable localization.