Multi-View and Multi-Scale Localization for Intelligent Vehicles in Underground Parking Lots

Abstract

Because of limited access to global positioning system (GPS) signals, accurate and reliable localization for intelligent vehicles in underground parking lots is still an open problem. This paper proposes a multi-view and multi-scale localization method aiming at solving this problem. The proposed method is divided into an offline mapping stage and an online localization stage. In the mapping stage, the offline map is generated by fusing 3-D information, WiFi features, visual features, and trajectory from visual odometry (VO). In the localization stage, WiFi fingerprint matching is exploited for coarse localization. Based on the result of coarse localization, multi-view localization is exploited for image-level localization. Finally, metric localization is exploited to refine the localization results. By applying this multi-scale strategy, it is possible to fuse WiFi localization and visual localization and reduce the image matching and error rate to a great extent. Because of exploiting more information, multi-view localization is more robust and accurate than single-view localization. The method is tested in a 2,000 m2 underground parking lot. The result demonstrates that this method can achieve sub-meter localization on average. The proposed localization method can be a supplement to the existing intelligent vehicle localization techniques.