Image similarity from feature-flow for keyframe detection in appearance-based SLAM

Abstract

In appearance based SLAM (Simultaneous Localisation and Mapping), a robot typically represents its environment through a set of acquired images that are associated with nodes in a topological map. Rather than storing every acquired image, which can be memory intensive, a selection of images (keyframes) representative of the places visited can be stored. Keyframe detection (i.e. choosing when to add a new keyframe) typically requires a means of determining the similarity of images. In this paper we develop three new metrics for computing image similarity. The metrics are based on the degree of feature-flow between features matched in a reference image (e.g. previous keyframe) and a test image (e.g. candidate keyframe), where a low degree of feature-flow indicates a high image similarity value. The new metrics and an existing metric are computed for synthetic and real data and their performance is evaluated with respect to a number of attributes important for keyframe detection. The results suggest similarity metrics based on feature-flow are preferable for use in keyframe detection.