A Coarse-to-Fine Method for Data-Efficient Collaborative Place Recognition

Multi-agent SLAM is a well-established and challenging problem in intelligent robotics. However, especially in decentralized systems, data usage is an important factor to consider because of the quadratic growth of peer connections and therefore data transmission. This paper proposes a method to reduce the necessary bandwidth for decentralized collaborative place recognition while preserving the accuracy of the place recognition matches. This method employs a preliminary, less data-intensive coarse evaluation based on a hierarchical bag-of-words model to exclude unlikely candidates. The method proposed in this paper is specifically designed for smaller groups of robots and shows comparable accuracy to the traditional approach for decentralized collaborative place recognition while reducing the necessary bandwidth by up to 40%. At the same time, our method shows promising results with regards to both accuracy and bandwidth requirements for groups of six or less robots compared to similar methods.