Visual Localization in a Prior 3D LiDAR Map Combining Points and Lines

Abstract

Robot localization in a prior map is critical for vision-based navigation. Currently, visual and visual-inertial odometry (VO&VIO) research has made great progress, but in the absence of closed loop, drift inevitably accumulates. The localization method using only 3D points had low accuracy and scale drift, and the localization algorithm using only straight lines had weak generalization ability and poor robustness. We propose a innovative and efficient visual localization in a 3D LiDAR map combining points and lines in this article. First, 3D lines are extracted offline from the 3D lidar map, and robust 2D lines in the raw image are extracted online. Prediction by Pose and reconstructed sparse 3D points from VIO or VO, we can efficiently obtain points and lines coarse correspondences respectively. The camera poses and correspondences are then iteratively optimized by minimizing the reprojection errors of 2D-3D lines and 3D-3D points. Our algorithm shows better results compared to state-of-the-art methods on the EuRoC MAV dataset, self-collection dataset and KITTI dataset without accumulated drifts.