Place recognition for visual loop-closures using similarities of object graphs

Abstract

A new method for object-based place recognition, where objects are used as a key intermediate representation of places, is proposed. Places are represented by object graphs and recognised from similarity scores between the generated graphs. An iterative Hungarian method is proposed for aligning graphs, where node and edge differences are used for measuring the similarity. The proposed technique robustly detects places without false positives, as it considers both the appearance and geometric information of places. Experiments are performed with mobile robots in indoor environments verified the effectiveness of the proposed method.