Lidar-Only Localization with 3D Pose-Feature Map

Abstract

Real-time and accuracy are the two most important indicators for unmanned vehicle localization. In this paper, we propose a novel map representation and its corresponding Lidar-only localization framework. In essence, we first extract geometric features from Lidar key-frames, and bundle these features with their observation poses (i.e., ground truth) to form a prior map named Pose-Feature Map. Then, the position of vehicle will be achieved by integrating Lidar-Odometry (LO) and Map-Matching (MM) with the Pose-Feature Map. In our framework, these two solutions are complementary. LO provides smooth and real-time pose estimation for MM, while MM can correct the accumulated drift of LO. During MM, we adaptively generate local maps to replace the global map for matching. Therefore, our proposed framework can further reduce the mismatch between the current frame and the map while maintaining low computational complexity. We demonstrated the framework in the KITTI dataset. The results confirm that our approach is superior to independent localization solutions in terms of real-time and accuracy.

This research is supported in part by National Key R&D Program of China under Grant 2019YFB2102400, in part by China Postdoctoral Science Foundation 2020M680906, in part by Hebei Province High-level Talent Funding project B202003027, in part by National Natural Science Foundation of China under Grant 61832013.