Lane marking based localization and mapping
A precise and robust map relative localization is necessary for our autonomous driving approaches. Since even high cost GNS systems suffer from shadowing effects and atmospheric disturbances, we use cameras to match a given map with our current field of view to compute the position of the vehicle.
A common visual feature on drivable roads is the presence of lane markings. Also, lane detection systems are widely used in vehicles.
In a first step we generate a map containing all visual lane markings. A preliminary version of the map can be obtained automatically. To eliminate misdetections or add additional information, we project the map on a virtual top-view image of the road and can manually verificate the data.
This top-view image can be obtained by using a velodyne laser scanner or camera systems combined with GNSS position information. In the latter case we compute disparity images from the stereo system and project the image to the ground plane.
In the online localization step, the measurements from a lane detection system are matched to the surrounding map area of the estimated position. In combination with data from an inertial navigation system (INS) the position is tracked with a Kalman filter.