Visual similarity analysis in loop closure through data dimensionality reduction via diffusion maps

Abstract

This paper describes loop closures detection, a significant problem in mobile robotics, using analysis of similarity between images in a low-dimensional mapping. We represent a set of images as a graph in high-dimensional space, where each node is represented by a dominant eigenvector of the correspondent image. To this graph, we apply Diffusion Maps by Coifman and Lafon [4], a graph-based spectral method to data dimensionality reduction. We determine visual similarity analysis and detect loop closure in lower dimension, without building a vocabulary of visual words. Our experiments show results of loop closures detection both in indoor and outdoor environments from images captured by an RGB camera as well as images captured using Google Street View.