LE BARZ Cédric

PhD graduated
Team : MLIA
Departure date : 06/30/2015
Supervision : Matthieu CORD
Co-supervision : HERBIN Stéphane

Visual autonomous navigation for small unmanned aerial vehicles

In this last decade, technology evolution has enabled the development of small and light UAV able to evolve in indoor and urban environments. In order to execute missions assigned to them, UAV must have a robust navigation system, including a precise egolocalization functionality within an absolute reference. We propose to solve this problem by mapping the latest images acquired with geo-referenced images, i.e. Google Streetview images.
In a first step, assuming that it is possible for a given query image to retrieve the geo-referenced image depicting the same scene, we study a solution, based on relative pose estimation between images, to refine the location. Then, to retrieve geo-referenced images corresponding to acquired images, we studied and proposed an hybrid method exploiting both visual and odometric information by defining an appropriate Hidden Markov Model (HMM), where states are geographical locations. The quality of achieved performances depending of visual similarities, we finally proposed an original solution based on a supervised metric learning solution. The solution measures similarities between the query images and geo-referenced images close to the putative position, thanks to distances learnt during a preliminary step.
Defence : 06/30/2015 - 10h - Site Jussieu 25-26/105
Jury members :
François Brémond, INRIA, [Rapporteur]
David Filliat, ENSTA, ParisTech [Rapporteur]
Matthieu Cord, Université Pierre et Marie Curie
Marcin Detyniecki, Université Pierre et Marie Curie
Jean-Yves Dufour, Thales
Stéphane Herbin, ONERA
Nicolas Thome, Université Pierre et Marie Curie
Ludovic Denoyer, Université Pierre et Marie Curie

2014-2015 Publications

 Mentions légales
Site map |