Lane identification based on robust visual odometry

Many important intelligent vehicle systems rely on lane-level position estimates. We propose a novel lane identification approach based on robust monocular visual odometry. The images obtained from the visual odometry camera as well as the trajectory estimate are used to construct a linearized representation of the world plane near the vehicle trajectory. This linear section is classified into road surface and non-road surface using a Gaussian Mixture Model. The width of the available road surface on either side is measured to detect extra drivable lanes. Coupled with road map annotations describing the number of lanes, this allows to determine the lane index of the vehicle. Preliminary experiments on a test set of 32 segments of 120m each prove the viability of the method.