Monocular Localization Within Manually Annotated LIDAR Maps

Vehicle localization is essential for autonomous driving in urban environments. Currently, most practical localization methods are based on laser scanners which are capable of measuring the distances precisely. However, multilayer LIDARs such as Velodyne HDL 64E are expensive and not applicable in consuming products. On the other hand, image sensor based localization relies on feature point matching and is sensitive to the change of weather and illumination. In this paper, we proposed a monocular vision based localization method within a prior map built with LIDAR. We manually label the lane markings in the LIDAR map during the offline stage. Then, during the testing time, the particle filter framework is employed to locate the vehicle with the visual cues. We extract the lane markings which are the salient features in the image and match them with the query results in the annotated map of each particle as the measurement. Experiments tested in the campus show that the proposed method can achieve lane-level localization.