Real-time SLAM using digital maps for complex and large scale urban environment

Abstract

To find out the location of the unmanned autonomous vehicles, accurate 3D map is one of the fundamental data to apply certain localization or matching algorithm[10, 53, 71, 77]. Recently, many research groups focus on city scale of reconstructing complex surroundings using mobile mapping system. In such a system, because the map is created based on the trajectory of the vehicle, it is the most important to calculate accurate position of the vehicle. Most vehicles utilize the Global Positioning System (GPS) sensor to estimate its location. However, in a complex urban environment, it is difficult to calculate accurate position using GPS sensor due to the blackout or multi-path propagation problem. Generally, using a variety of sensors is suggested in order to complement the drawbacks of a single sensor. This thesis discusses an approach for the large-scale 3D mapping of complex urban environment using Simultaneous Localization and Mapping (SLAM) algorithm which uses the matching of digital map and point cloud. First, we have built a mobile mapping system suitable for urban environment that includes Light Detection and Ranging (LiDAR), GPS, Inertial Measurement Unit (IMU), camera, wheel encoder and altimeter sensors. To achieve a dense map of the urban area, we introduce a local point cloud generated by two 2D LiDAR[45]. Second, we seek solutions to challenges in SLAM-based pose estimation and 3D reconstruction as addressed by [29, 47]. To be used in SLAM, digital maps, such as BI(Building Information), AI(aerial image)and Digital Elevation Model (DEM), need intelligent representation [62]. Using digital map available publicly, we propose a matching based approach that uses digital map and accumulated local point cloud to correct the heading and translation error from dead reckoning navigation. The proposed heading correction algorithm is validated using the publicly available KITTI dataset. Third, we present intelligent map representation for unmanned autonomous vehicles and exploitation of digital maps. Compact and useful representation is essential. Raw 3D point cloud data, because it contains much unnecessary information, we apply probabilistic voxel for abstraction of point cloud, to achieve efficiency in terms of memory and computation time. Results are present for large scale city reconstruction experiments using the two Urban Mapping Systems. Our real-time map generation framework is validated via a long-distance urban test and evaluated at randomly sampled points using RTK-GPS. For thorough validation, we ran the algorithm on three levels of urban areas with di erent GPS availability and structural complexity

a downtown with mid-complexity, a building complex and the downtown area of a large city.