Scene recognition with omnidirectional images in low-textured environments

Abstract

A combined method involving global and local descriptors was developed to recognise scenes for loop closure detection in low-textured environments. An omnidirectional image is divided into background regions and salient regions according to the colour distribution. To represent a scene with features that are appropriate to its characteristics, global features for background regions are calculated and scale invariant feature transform features for salient regions are extracted. The proposed method can compute a more distinct scene similarity, and this was verified by an experiment involving loop closure detection.