Towards an accurate topological localization using a Bag-of-SIFT-visual-Words model

Abstract

Topological localization is a problem in mobile robotics that implies the ability of an agent to self locate in an environment. In this paper, we approach the task of topological localization without using a temporal continuity of the images of the places the robot has been. The environment is represented by an office under different illumination settings acquired with a perspective camera mounted on a robot platform. We create visual vocabularies based on invariant local features and different distance-based K-means clustering. The experimental setup is performed with an One-versus-All classifier with different kernel functions that achieved success.