In this thesis, we introduce a novel map representation that combines object identity with metric information for convenient human-robot interaction.
Previous point-based maps provide good metric information. This map representation, however, is illegible for human users because it is hard to conceive a characteristic object from a cloud of points. And the point-based representation, necessarily, involves in an exhaustive Markov search for localization due to the weak distinctiveness of the points. Moreover the number of landmarks in the map increases linearly even when a robot builds a map in a large scale environment. This causes the scalability problem in SLAM.
To solve these problems, we propose a novel hierarchical SLAM algorithm that produces a map that contains both object identities and metric information. This map representation is not only easy for human users to read, but also provides better distinctiveness by using the objects as landmarks. By using the object representation coupled with metric information, the number of landmarks also drastically decreases.
We analyze the proposed algorithm through various experiments. The algorithm shows good results in both mapping and localization. The kidnap problem is also directly solved without a markov search due to the distinctive nature of the representation.