On LIDAR Map Combination: A Graph Slam Module to Generate Accurate and Largescale Maps for Autonomous Driving

Precise map combination is the key solution to generate largescale maps for autonomous vehicles. The current state-of-art of LIDAR based SLAM technologies does not allow to achieve this demand because of dealing with a huge number of vehicle positions and the sparsity of point clouds to represent environments. In this paper, we fully design a unique Graph Slam framework to combine maps based on node strategy. A node encodes a set of LIDAR frames and represents accurate road surfaces in a grayscale image. The nodes images are identified in the Absolute Coordinate System (ACS) to facilitate the detection of map-combiner events between maps. The relative-position errors at the detected map-combiner events are significantly compensated by matching the dense nodes' images using the Phase Correlation technique. These compensations are reflected in ACS by a cost function that is designed to combine maps with maintaining the consistency of the road surfaces in the image domain. To prove the reliability of the proposed system, the world's longest tunnel has been scanned two times. In addition, two combination formulas are investigated to simulate the creation and updating of collected maps by different agents. The experimental results have verified the robustness, accuracy and outperformance of the proposed framework against an accurate GNSS/INS-RTK system with improving the global position accuracy.