Generation of Accurate Lane-Level Maps from Coarse Prior Maps and Lidar

While many research projects on autonomous driving and advanced driver support systems make heavy use of highly accurate maps covering large areas, there is relatively little work on methods for automatically generating such maps. These maps require accuracy in both the number of lanes and positioning of every lane, which we call lanelevel maps. Here, we present a method that combines coarse, inaccurate prior maps from Open-StreetMap (OSM) with local sensor information from 3D Lidar and a positioning system. We formulate a probabilistic model of lane structure using such information, and develop a number of tractable inference algorithms. These algorithms leverage the coarse structural information present in OSM, and integrates it with the highly accurate local sensor measurements. The resulting maps have extremely good alignment with manually constructed baseline maps generated for autonomous driving experiments.