Robust Localization Using 3D NDT Matching and Beam Model for Autonomous Vehicles in an Urban Scenario with Dynamic Obstacles

Accurate and robust positioning is significant for autonomous vehicles. Generally, the LiDAR-based map matching (LMM) can render lane-level localization accuracy by matching 3D real-time point clouds with a pre-built point cloud map. However, both the robustness and accuracy of matching are deteriorated in the environments with excessive dynamic obstacles, such as moving vehicles, occasionally blocking the pre-built map. The dynamic obstacles are not contained in the pre-built map. The major solution to cope with the excessive dynamic objects is to detect and remove them from the real-time point clouds. However, the performance of this solution relies on the accuracy of object detection and requires high computation power. To improve the performance of LiDAR-based positioning in dynamic environments, this paper proposes to incorporate two different LiDAR measurement models, the 3D normal distribution transform (NDT) and beam model. This paper takes the complementary properties of both the 3D NDT and beam model to improve
the performance of LiDAR-based localization in dynamic environments. The proposed method is verified through real road tests in dynamic road of Berkeley, California, USA. According to the experiment results, the proposed method outperforms the standalone NDT matching.