Loop closure detection with a holistic image feature

Abstract

In this paper we introduce a novel image descriptor, LBP-gist, suitable for real time loop closure detection. As the name suggests, the proposed method builds on two popular image analysis techniques: the gist feature, which has been used in holistic scene description and the LBP operator, originally designed for texture classification. The combination of the two methods gives rise to a very fast computing feature which is shown to be competitive to the state-of-the-art loop closure detection. Fast image search is achieved via Winner Take All Hashing, a simple method for image retrieval that exploits the descriptive power of rank-correlation measures. Two modifications of this method are proposed, to improve its selectivity. The performance of LBP-gist and the hashing strategy is demonstrated on two outdoor datasets. © 2013 Springer-Verlag.