Building Visual Maps From Two-Camera Views for Intelligent Vehicle Localization in Underground Parking Lots

Robust and reliable localization is important for intelligent vehicles, especially in GPS-blind areas, such as underground parking lots. A novel localization frame by building two-layer visual maps from two-camera views (i.e., front and side views) is proposed in this paper. Among these two views, we build the dense map layer by extracting handcraft features from the front-view to achieve accurate localization. We build the sparse map layer by extracting both holistic and local semantic features with a Siamese Network from the side-view to improve localization reliability. These two types of map layers together with the 3D information are integrated to formulate the visual map. In the localization stage, we propose a map-based localization scheme, and a novel method to fuse the holistic and local semantic features matching results. By two-view matching within the pre-built visual map, we can measure the poses of the vehicle and achieve both accurate and reliable localization. The proposed mapping and localization methods have been tested in two experimental scenes about 3500 m² and 2000 m². Experimental results demonstrate that the final localization has a better performance after using the proposed method.