A Robust LiDAR State Estimation and Map Building Approach for Urban Road

Abstract

In a dynamic environment, the long-term and reliable positioning of the robot system is the key. Good automatic driving needs to solve the influence of multiple dynamic objects in the road environment, which puts forward higher requirements for the location accuracy and the robustness of the self-driving vehicle. In this article, we mainly research the positioning problem in the urban road environment, and we design a lidar location and navigation system based on multi-sensor fusion. First, the low-cost VLP-16 lidar is used as the front end of the system to obtain the surrounding high-precision point cloud, and the inertial measurement unit (IMU) is used to replace the uniform motion model to remove the movement distortion of the lidar. The optimized point cloud information constructs a local map. After that, we combined low-cost GPS information to optimize the global map accuracy by using map optimization. Finally, the local map saved on the hard disk is generated through our optimization method to generate a global laser point cloud map of any size. At the same time, we found in actual tests that a major factor that affects the positioning accuracy of lidar on urban roads is the dynamic and semi-static moving vehicles on the road. In order to not be affected by the occlusion of parallel vehicles on the front and rear of the road when building the map, the deep point cloud neural network is used. The network recognizes these vehicles, and extracts point clouds through multi-target tracking, and removes the corresponding point clouds in the local and global maps. Autonomous driving vehicles in the park that meet the above test requirements are tested. The experimental results show that the multi-sensor fusion and dynamic target removal positioning system we designed meets the requirements of urban road positioning under multiple dynamic targets.