Abstract:

To enhance the localization performance of intelligent vehicles using light detection and ranging (LiDAR), the concept of point cloud fingerprints is introduced, along with a map representation model and a localization method based on point cloud fingerprints. The point cloud fingerprint representation relies on point cloud polarization and principal component features, integrating global and local descriptions of the point cloud. The map representation model consists of sequential nodes containing only point cloud fingerprints and global poses. The multiscale localization of intelligent vehicles based on point cloud fingerprint maps is decoupled into three modules: coarse localization based on standard global navigation satellite system (GNSS) and trajectory prediction, node localization based on point cloud fingerprints and the vector-based Pearson correlation coefficient (RV), and metric localization based on point cloud fingerprints and the fast generalized iterative closest point (Fast-GICP). Experiments conducted on the publicly available KITTI dataset and a local dataset demonstrate that the proposed method achieves over 99% map-matching accuracy and up to 7 cm localization accuracy. The results indicate that the proposed method attains centimeter-level high-precision localization in various scenarios, with strong robustness and generalization capabilities.

Key words: point cloud fingerprint, polarization representation, principal component feature, map representation, multiscale localization