| Abstract: | Accurate vehicle self-localization is significant for autonomous driving. The localization techniques based on Global Navigation Satellite System (GNSS) cannot achieve the required accuracy in urban canyons. On the other hand, simultaneous localization and mapping (SLAM) methods suffer from the error accumulation problem. State-of-the-art localization approaches adopt 3D Light Detection and Ranging (Lidar) to observe the surrounding environment and match the observation with a priori known 3D point cloud map for estimating the position of the vehicle within the map. However, storing the massive point cloud needs immense storage on the vehicle, or it should be stored on servers, which makes the simultaneous downloading of the map by multiple vehicles another challenge. In this study, rather than employing the point cloud directly as the prior map, we focus on the abstract map of buildings, which is easy to extract, and at the same time apparently observable by Lidar. More especially, we proposed vehicle localization methods based on two different abstract map formats representing urban areas. The first format is the multilayer 2D vector map of building footprints, which represents the building boundaries using vectors (lines). The second format is the planar surface map of buildings and ground. These map formats share the same idea that the uncertainty (deviation) of each vector or planar surface is calculated and included in the map. Later in the localization phase, the observed data from Lidar is matched with the abstract map to obtain the precise location of the vehicle. Experiments conducted in a dense urban area of Tokyo show that even though we significantly shrank the map size, we could preserve the mean error of the localization. |
