Lane markings-based relocalization on highway

In this paper we propose a method to create maps of lane markers and a relocalization method using these maps on highways. We use a three-steps sub-pixellic detection process to locate the edges of lane markers with centimetric precision using images from a front-facing camera. We propose a simple mapping process using RTK-GNSS data to create lane marker edge maps to be used in the relocalization task. Given such a map, we then use transferable belief model to associate observed markers to those stored in the prior map, and a symbolic registration algorithm is then used to estimate the raw pose of the vehicle. Finally, this raw pose measurement is fused with odometric data to improve the global pose estimate precision. This method is evaluated on an open highway driving sequence of 10km on the outskirts of Paris. The proposed approach works in real time and yields lateral and longitudinal errors with standard deviations of respectively 8 and 13 centimeters.