The Method and Performance Analysis of Constructing Visual Point Cloud Map to Assist Inertial Positioning

摘要

In the face of discontinuous GNSS signal and INS error accumulation in complex urban environment, the research on multi-source information fusion positioning method assisted by high-precision map is essential. In order to meet the real-time, continuous and reliable positioning requirements of vehicle navigation, this paper presents the principle and method of constructing high-precision map with stereo vision, and introduces the key technologies of map generation and data cleaning. Meanwhile, it analyzes the performance of the map-aided visual-inertial fusion positioning method. Firstly, the visual point cloud map is constructed based on mobile surveying. The local coordinates of road marking points are obtained by the front intersection of common-view pictures, then the camera pose solved from GNSS/SINS converted road marking points from local coordinate system to ECEF. However, the quantity of road marking points is large, and the absolute precision is difficult to evaluate. Therefore, this paper gives the quality indexes of evaluating point cloud map precision and data cleaning method. By using the method of deep learning to target recognition and semantic segmentation, only the long-term static stable road markings can be maintained in the map. In the meantime, a visual positioning method based on Octomap and DBOW acceleration is proposed. The K1TTI dataset test shows that point cloud map is constructed and an average data cleaning rate is 35.05%, that many error observations are excluded. When it comes to positioning, the accuracy of visual/GNSS/INS integrated positioning yields about 2 cm and 0.06°. When integrated with INS, a positioning accuracy of 10 cm can be maintained with only one road landmark matched successfully in the case of GNSS long time (20 min) unlock.