Scan Context 3D Lidar Inertial Odometry via Iterated ESKF and Incremental K-Dimensional Tree

Abstract

This thesis focused on a 3D lidar inertial odometry algorithm framework that improves the Lightweight and ground optimized lidar odometry and mapping (LeGO-LOAM) by constructing a new back-end optimization algorithm. In comparison with the LeGO-LOAM, the feature extraction and image projection processes are still the same. Two step Levenberg Marquardt method was replaced with an iterated ESKF method in the lidar odometry to produce a better initial pose for the robots, and the k-dimensional(k-d) tree method in the lidar mapping is replaced with the ikd-Tree method to ensure high performance mapping process in real time. In the loop closure, a scan context search method is added to better correct the algorithm’s final trajectory.

The proposed algorithm is tested and simulated by configuring the robot operating system (ROS) with the ubuntu virtual Linux system. The performance of the optimized back-end algorithm has compared with other three algorithms to show the proposed algorithm has better accuracy in the localization and mapping process.