Probabilistic approach to self-localization for autonomous mobile outdoor robotics based on hybrid map knowledge

We propose a probabilistic approach to self-localization of an autonomous mobile outdoor robot based on hybrid (metrically-annotated topological) map knowledge. A particle filter is used to estimate the state of the robot based on map data and relative dead-reckoning measurements. Convergence speed of the filter is improved by equidistant road sampling placing particles equidistantly along map segments. We introduce a novel particle weighting technique which realizes a comparison of a local vehicle motion buffer to local map data to increase the robustness of the estimation process by significantly reducing the occurrence of ambiguous situations during the localization process. The weighting technique is combined with a custom resampling method generating an increased attachment of particles to nearby road segments. We present experimental results of long-distance test runs featuring various difficult local map configurations demonstrating the robustness and quality of the proposed algorithm even in presence of map errors.