Network cameras
MSc dissertation proposal 2009/2010
Network Cameras
Calibration
Introduction:
The increasing need of surveillance in public spaces and the recent
technological advances on embedded video compression and communications made camera
networks ubiquitous. Typical environments include single rooms, complete
buildings, streets, highways, tunnels, etc. While the technological advances already
allowed such a wide installation of camera networks, the automatic
understanding and processing of the video streams is still an active research area.
One of the crucial problems in camera networks is to obtain a correct
calibration of each camera in terms of a unique reference frame. This condition
is a fundamental feature required for further higher level processing (i.e.
people/car tracking, event detection, metrology) and nowadays one of the most active
research subjects in Computer Vision. The problematics
generally arise from the lack of overlapping field of views (FOV) of the camera
which does not allow the estimation of a common reference frame for each
camera. In such scenario, exactly geolocating each
sensor is extremely complex without the aid of special equipment (moving
calibration patterns or GPS) or a priori reference images such as a panorama of
the given environment.
Objectives:
The main objective of the work consists in estimating the calibration of
a network of cameras, possibly with non-overlapping fields of view. Calibration
comprises both the intrinsic and extrinsic parameters of the cameras. The
non-overlapping problem is expected to be overcome using a mobile robot with
the capability of estimating of its pose in a global frame. The robot is
equipped with one calibrated camera which we assume that can be oriented in a
manner to observe world points also seen by the network of cameras.
Detailed description:
In order to calibrate the camera network we propose an approach based on
the visual reconstruction of some points of the scenario. These points are
expressed in a global (world) coordinate frame provided by self-localization
information assumed to exist in a mobile robot. Given the reconstructed
3D-points of the environment and their images we can calibrate the network of
cameras using standard computer vision methodologies [Hartley00]. The
reconstruction of 3D points comprises two main steps, namely matching image
points and computing their locations. We do the matching based on Scale
Invariant Feature Transform (SIFT) features, state of the art features well
known to provide a very robust matching procedure [LoweWWW,
Lowe04], and the computation of the 3D locations is based on Visual
Simultaneous Localization and Map Building (vSLAM)
[Goncalves05, Karlsson05].
Work-proposal detailed steps:
(i) Studying and testing vSLAM,
by doing the reconstruction of points and trajectories in some simulated or
real scenarios, based on various software packages freely available in the
Internet.
(ii) Testing the detection and correspondence of SIFT in simulated or
real scenarios. Studying some methodologies to improve the precision of the
registration of the SIFT features.
(iii) Testing the camera calibration methodologies based on the
information obtained in the previous steps.
References:
[Leite08] Calibrating a Network of Cameras Based
on Visual Odometry, Nuno
Leite, Alessio Del Bue, José Gaspar, in Proc. of IV Jornadas de
Engenharia Electrónica e Telecomunicações e de Computadores, pp174-179, November 2008, Lisbon, Portugal.
[Goncalves05] L. Goncalves, E. Di Bernardo, D.
Benson, M. Svedman, J. Ostrowski,
N. Karlsson, and P. Pirjanian.
A visual front-end for simultaneous localization and mapping. In International
Conference on Robotics and Automation, 2005.
[Karlsson05] N. Karlsson, E. Di Bernardo, J. Ostrowski, L. Goncalves, P. Pirjanian, and M. Munich. The vslam
algorithm for robust localization and mapping. In International Conference on
Robotics and Automation, 2005.
[Hartley00] R. I. Hartley and
[LoweWWW] David G. Lowe. Demo software: Sift keypoint detector. http://www.cs.ubc.ca/˜lowe/keypoints/.
[Lowe04] David G. Lowe. Distinctive image features from scale-invariant keypoints. In International Journal of Computer Vision,
pages 91–110, 2004.
Requirements (grades, required courses, etc):
-
Expected results:
At the end of the work, the students will have enriched their experience
in computer vision applied to camera network setups. In particular are expected
to develop and assess:
- precise image (feature) registration methodologies
- precise calibration methodologies
Place for conducting the work-proposal:
ISR / IST
More MSc dissertation
proposals on Computer and Robot Vision in:
http://omni.isr.ist.utl.pt/~jag