Observable Point Cloud Maps for Wide-Area Self-Localization

Self-localization using a pre-constructed point cloud map is effective even in global navigation satellite system-denied settings, such as indoor environments and areas adjacent to buildings. In wide-area environments, however, the amount of map data becomes enormous, increasing storage and computational costs. This study proposes a method to partition a wide-area point cloud map into local regions and construct observable point cloud maps by extracting only the points observable from virtual viewpoints placed within each region. Each observable point cloud map contains only the point cloud that contributes to self-localization in the corresponding local region. During the self-localization process, the robot dynamically switches the observable point cloud maps according to its current local region. Self-localization accuracy and computational costs are evaluated using both the wide-area point cloud map and proposed observable point cloud maps. Additionally, the localization contribution per point is used to quantify the effectiveness of the retained map points. The results show the dynamic switching of observable point cloud maps enables memory-efficient self-localization while maintaining localization accuracy comparable to that of the wide-area point cloud map.