Accurate localization is one of the fundamental tasks of vehicles visual navigation in parking lots. In this paper, we propose a practical and novel solution, which exploits road marking semantic segmentation to attack the problem of long-term and high-precision visual localization. Based on the semantic data association derived from road markings segmentation, point cloud fusion and loop detection strategies are designed to improve the performance of semantic map building. Applying the generated map, we present a point cloud registration algorithm combining semantic and geometric inference to improve the localization precision. Experiments on real-world indoor parking lots prove that the semantic map created by the proposed method reveals more accurate and consistent performance. Moreover, localization error is no more than 10cm, while running in real-time performance.
Mapping and Localization using Semantic Road Marking with Centimeter-level Accuracy in Indoor Parking Lots
2019-10-01
1058814 byte
Conference paper
Electronic Resource
English