Inc-DLOM: Incremental Direct LiDAR Odometry and Mapping

Abstract

Intelligent Vehicle (IV) research is gaining popularity due to the convergence of technological advancements and societal demands, which also leads to the fundamental demand for precise localization. However, the localization accuracy of most existing LiDAR Odometry methods is limited by the complex environment and high-frequency motion, leading to unsatisfactory performance. Moreover, the point cloud data generated by different LiDARs will possess different properties, such as spatial density, Field-of-View (FoV), perception distances, etc., which may have a great impact on LO methods, and makes the generalization of LO a noteworthy issue. To address these issues, we propose the method of Incremental Direct LiDAR Odometry and Mapping (Inc-DLOM). Our proposed Inc-DLOM has the following key contributions: 1) a voxel-to-voxel (V2V) scan matching scheme for scan-to-scan transform estimation; 2) the Incremental Voxel Mapping (IVM) method to incrementally update and maintain the historical mapping information; 3) the Incremental GICP solver to refine the global pose by IVM. To evaluate the performance in terms of accuracy and efficiency, extensive experiments have been conducted with both mechanical LiDAR and solid-state LiDAR on different robotic platforms, including public datasets and real robot data acquisition. The experimental results show that Inc-DLOM achieves better accuracy, efficiency, and generalizability than other comparison state-of-the-art LiDAR Odometry methods.

Keywords

• Autonomous driving
• LiDAR odometry
• SLAM
• 3D point cloud