Detection and tracking of rural crossroads combining vision and LiDAR measurements

Perceiving road networks plays a key role in navigating a robot on the road. The estimation of the robot´s ego lane grows in complexity if crossroads have to be passed. Especially in rural areas, where crossroad locations are more likely to be erroneous, perceiving them in sensor data can prevent improper turn maneuvers. Vice versa the recognition of intersections can be used for localization a given road map. Most of the proposed methods for intersection detection use single sensors like camera or LiDAR. In this paper we show a novel approach for detecting and tracking of rural crossroads by fusing multiple sensors in an accumulated terrain map. Further information for each cell within the terrain map is calculated by extracting the road area and correlating it with different road geometries. In order to detect intersections, we build several hypothetical intersection centers and emit rays from their position. The rays are evaluated according to their cells´ information and are bundled to build the branches of intersections. A Sequential Extended Kalman Filter estimates position and geometry of the detected intersections over time. Smart data assignment allows to partially correct single branches of a crossroad as well as to add new branches. The algorithm´s robust performance under several challenging conditions can be seen at http://www.mucar3.de/itsc2014.