Thesis Defense: Towards Lifelong Visual Localization and Mapping


Event Speaker: 

Presenter: Hordur Johannsson

Event Location: 

32-G449 (Kiva)

Event Date/Time: 

Thursday, January 31, 2013 - 10:00am

Lifelong autonomy for mobile robotic systems requires algorithms that
are robust and scale efficiently with time as sensor information is
continually collected. For mobile robots one of the fundamental problems
is navigation; which requires the robot to have a map of its
environment, so it can plan its path and execute it. Having the robot
use its perception sensor to do simultaneous localization and mapping
(SLAM) is beneficial for a fully autonomous system. Extending the time
horizon of operations poses problems to current SLAM algorithms, both in
terms of robustness and temporal scalability. To address this problem we
propose a reduced pose graph model that scales with the size of the
environment instead of time. Additionally we develop a SLAM system using
two different sensors modalities: imaging sonars for underwater
navigation and vision based SLAM for terrestrial applications.

Underwater navigation is one application domain that benefits from SLAM,
where access to a global positioning system (GPS) is not possible. In
this thesis we present SLAM systems for two underwater applications.
First, we describe our implementation of real-time imaging-sonar aided
navigation applied to in-situ autonomous ship hull inspection using the
hovering autonomous underwater vehicle (HAUV). In addition we present an
architecture that enables the fusion of information of a sonar and a
camera system. The system is evaluated on data collected during
experiments on SS Curtiss and USCGC Seneca. Second, we develop a feature
based navigation system supporting multi-session mapping and provide an
algorithm for re-localizing the vehicle between missions. In addition we
present a method for managing the complexity of the estimation problem
as new information is received. The system is demonstrated using data
collected with a REMUS vehicle equipped with a BlueView forward looking

The model we use for mapping builds on the pose graph representation
which has been shown to be an efficient and accurate approach to SLAM.
One of the problems with the pose graph formulation is that the state
space continuously grows as more information is acquired. To address
this problem we propose the reduced pose graph (RPG) model which
partitions the space to be mapped and uses these partitions to reduce
the number of poses used for the estimation. To evaluate our
approach, we present results using an online binocular and RGB-D
visual SLAM system that uses place recognition for both robustness and
multi-session operations. Additionally, to enable large-scale indoor
mapping, our system automatically detects elevator rides based on
accelerometer data. We demonstrate long-term mapping using
approximately nine hours of data collected in the Stata Center over
the course of six months. Ground truth, derived by aligning laser
scans to existing floor plans, is used to evaluate the global accuracy
of the system. Our results illustrate the capability of our visual
SLAM system to scale in size with the area of exploration instead of
the time of exploration.

Chair of the defense: Prof. Franz Hover
Thesis Committee: Prof. John Leonard, MIT, advisor
Prof. Seth Teller, MIT
Prof. Daniela Rus, MIT
Dr. Hanumant Singh, WHOI
Dr. Michael Kaess, MIT

Presenter's Affiliation: CSAIL, MIT/WHOI JP