부산대학교 도서관

One of the huge challenges of map-based localization is a rapidly changing environment. The present contribution addresses this problem by first constructing a new framework for feature-based long-term mapping using a Bernoulli filter. This framework is then applied to construct a continuously learning map. It is based on Simultaneous Localization and Mapping (SLAM) to create a short-term map which provides a momentary image of the environment during one mapping run. The proposed fusion algorithm then estimates the landmark map on a long-term basis by incorporating those short-term maps. Landmarks in the long-term map that reach a negligible spatial uncertainty can then be used again as a prior for the short-term mapping process. Since a Bernoulli filter can only handle a single object, independent groups of landmarks are constructed where only those with exactly one landmark are updated. As a result, landmarks that are part of the long-term map are quite distinct. By incorporating additional but probably outdated a priori information, the proposed method is able to restrict the inevitable error propagation of SLAM algorithms. The long-term mapping process is further distributable to several agents: every agent simultaneously localizes itself while it generates a new snapshot that is fused into the long-term map afterwards. An evaluation using real-world data completes this contribution.