[robotics-worldwide] PhD position visual slam

Youcef Mezouar mezouar at univ-bpclermont.fr
Wed Sep 29 05:39:53 PDT 2010

The Lasmea (Joint unit of CNRS/ Blaise-Pascal University, France) and 
ISRC/SKKU (Intelligent System Research Center of  Sungkyunkwan 
University, Korea) are offering 1 PhD grant (duration: 3 years). This 
position is opened in the framework of the French/Korean BRI project 
“Cognitive Personal Transport Service Robot”  sponsored by the 
Gyeonggi-province International Collaboration Research Project. The 
student is expected to spend 2 years in Clermont-Ferrand, France and 1 
year in Korea.

Good mathematical background
basic knowledge on computer vision, control and robotics
Programming (C++, Matlab).

To apply, please send CV, the Master results, a statement of interest to

Youcef MEZOUAR: Youcef.Mezouar at lasmea.univ-bpclermont.fr
Home page: http://gravir.univ-bpclermont.fr/personne/Youcef.Mezouar/

More details about the PhD subject:

Curent visual Simultaneous Localisation An Mapping (SLAM or vSLAM) 
algorithms have some shortcomings when they are applied to large scale 
environments. First the algorithmic complexity grows quadratically with 
the size of the map, and large scale maps are intractable. Moreover, 
maps generated with SLAM or Structure From Motion (SfM) algorithms have 
some large scale drift coming from error accumulation.

However, in the case of mobile robot navigation, local maps are 
sufficient to allow autonomous navigation. Large scale drift doesn't 
matter because the robot sensors can only see nearby landmarks. On the 
other hand, for tasks such as path planning, using the whole map is 
mandatory. The same is true for matching locations between the SLAM map 
and a Geographical Information System or an aerial image. In those 
cases, the map needs to be topologically correct. That means the 
transformation between the map and the real world should be bijective 
and bicontinuous. To achieve that, it is necessary to solve the well 
known problem of the loop closing.

The objective of this thesis will be to propose a SLAM system able to 
deal with the above mentioned difficulties. Two main problems have to be 
adressed. First the representation of the map should be able to decouple 
the local geometry from large scale structure and topology. The map 
should be able to accept large scale deformations or loop closures with 
a low computational cost and without modifying the local geometry. The 
second problem to solve is the detection of loop closures. This can be 
done by combining 2D photometric information from images and 3D 
information from the local geometry.

More information about the robotics-worldwide mailing list