Three level sequence-based Loop Closure Detection

Abstract

The recognition of previously visited places, known as Loop Closure Detection (LCD), composes one of the problems widely studied in robotics: simultaneous localization and mapping (SLAM). In this paper we propose a three level hierarchy based LCD method. In our serialized approach, in the First Level, a sequence of the most recently visited places is used as query to search for candidate sequences in our topological map composed by previously visited places. After that, at the Second Level, the method selects the most similar sequence to the query among all candidate sequences which is temporally consistent with the previous LCD method response. Then, at the Third Level, we match the image sequences belonging to the query sequence to the candidate sequence selected in the Second Level. The method is evaluated in different and challenging public datasets, and presents expressive results that overcome the LCD state-of-the-art methods.