Improving Vehicle Localization Using Semantic And Pole-Like Landmarks

In this paper, we present a framework for vehicle self-localization in urban environments. It utilizes semantic and distinctive physical objects such as trees, traffic signs or street lamps as robust landmarks and deduces the global vehicle pose in conjunction with an offline map. Since it is independent from the availability of road markings and the knowledge of street courses, application in dense urban areas with high rates dynamic objects and road users is possible. This paper introduces novel methods for vehicular environment perception via LiDAR scanner and stereo camera, as well as models for their association with a high-precision digital map to estimate the vehicle's position via Adaptive Monte-Carlo Localization. Evaluation in urban areas indicates the potential for global positioning accuracy below 0.30 m for LiDAR and below 0.50 m for stereo camera, as well as a corresponding heading error below 1 degrees.