Robust offline topological map estimation using visual loop closures

Abstract

A framework employing the Student-t pdf is introduced for offline map estimation and robot localization using visual loop closures. The framework uses the Student-t pdf (a) as an observation model of a Hidden Markov Model to represent a topological map (b) to represent the robot motion model. The map and the motion model are calculated in an expectation maximization (EM) framework. We show that the estimator converges at linear time and that the provided accuracy is higher compared to using a conventional Gaussian mixture pdf, due to higher noise resiliency, as well as compared to using a fixed robot motion model. The task is assisted by unsupervised landmark definition through the EM-based clustering of the observations and by scene representation using the complex Zernike moments, which provide rich rotation-invariant information. The validity of the method has been verified experimentally using the input from an omnidirectional camera.