On Ambiguities in Feature-Based Vehicle Localization and their A Priori Detection in Maps

A major risk to reliable and accurate feature-based vehicle localization lies in the data association. Data association is the process of matching localization features that are extracted from the current sensor data with the corresponding landmarks in a high-definition map. Due to similar and repetitive structures in the environment, matching sensor data with map data becomes highly ambiguous. These ambiguities are difficult to detect online, jeopardize correct and accurate localization, and risk localization integrity. In this paper, the existence of ambiguities arising from similar constellations of localization features is evaluated. We propose a novel Geometric Hashing based method for the detection of such patterns in maps which enables an a priori estimation of expected ambiguities during localization. Our approach is demonstrated by the evaluation of two map sections covering an urban loop of 4 km as well as a route of 40 km on a German highway. Both maps contain cylindrical objects such as signs, delineators, and trees for localization. We show that many ambiguous patterns of various sizes exist, extract all of their occurrences in both maps, and derive statistics on their distribution in the map frame.