Invariant Spatial Information for Loop-Closure Detection

Recently, Bag-of-Words (BoW) has become a de-facto standard solution for loop-closure detection (LCD) in robotic visual SLAM (rvSLAM). Whereas BoW is efficient in using appearance information as invariant feature for comparing query and mapped scenes, it is not straightforward to use spatial information as invariant feature in BoW. In this paper, we propose a new LCD approach, termed invariant spatial information (ISI), which transforms spatial information in every query/mapped image into an invariant coordinate system by exploiting the available 3D map in rvSLAM. This enables direct comparison of feature location between different images, and allows us to use both appearance and spatial information as invariant feature. Experiments show that the proposed ISI-based LCD outperforms existing LCD methods.