Topological place recognition based on long-term memory retrieval

Topological place recognition is related to the retrieval of previously learned places from long-term memory. In this paper, we consider this problem and present a novel approach - based on the previously proposed bubble descriptor semantic tree (BDST) memory model. In the proposed approach, the robot combines decision-making at each searched node of the BDST along with a BDST traversal strategy in order to find the most related previous knowledge. In case the robot is kidnapped or has no knowledge of where it is coming from, the traversal uses top-down depth-first search. If the robot has been navigating and knows where it is coming from, it uses this knowledge to initiate its search in an integrated bottom-up and top-down manner. The experimental results indicate that the proposed approach generally improves recognition performance significantly in comparison to purely top-down traversal.