WiCRF: Weighted Bimodal Constrained LiDAR Odometry and Mapping With Robust Features

Accurate localization is a fundamental capability of autonomous driving systems, and LiDAR has been widely used for localization systems in recent years due to its high reliability and accuracy. In this paper, we propose a robust and accurate LiDAR SLAM, which innovates feature point extraction and motion constraint construction. For feature extraction, the proposed adaptive point roughness evaluation based on geometric scaling effectively improves the stability and accuracy of feature points (plane, line). Then, outliers are removed with a dynamic threshold filter, which improves the accuracy of outlier recognition. For motion constraint construction, the proposed weighted bimodal least squares is employed to optimize the relative pose between current frame and point map. The map stores both 3D coordinates and vectors (principal or normal vectors). Using vectors in current frame and point map, bimodal reprojection constraints are constructed. And all constraints are weighted according to the neighboring vector distribution in the map, which effectively reduces the negative impact of vector errors on registration. Our solution is tested in multiple datasets and achieve better performance in terms of accuracy and robustness.

Automated summary

Accurate localization is crucial for autonomous driving systems, and LiDAR is commonly used due to its reliability. This paper proposes a robust and accurate LiDAR SLAM that innovates feature point extraction and motion constraint construction. Feature points are extracted using adaptive point roughness evaluation and outliers are removed with a dynamic threshold filter. Motion constraint construction uses weighted bimodal least squares to optimize the relative pose between current frame and point map. The solution achieves better performance in terms of accuracy and robustness according to multiple datasets.