Background difference example
MSc dissertation
proposal 2007/2008
Target Detection using Adaptive Background Models
Introduction:
"Although surveillance cameras are already prevalent in banks,
stores, and parking lots, video data currently is used only -after the fact- as
a forensic tool, thus losing its primary benefit as an active, real-time
medium. (...)"
[introduction of the Video Surveillance And
Monitoring project - http://www.cs.cmu.edu/~vsam/index.html]
Today we assist to cameras installed
everywhere; however we find that most of them are unattended due to the very
high costs of having humans watching the huge amounts of the resulting
video-imaging. In this work we propose building an automatic target detection method,
in order to minimize the surveillance work done by humans, and in addition
using pan-tilt-zoom (PTZ) cameras which allow simultaneously covering large
surveillance areas and observing regions of interest in detail.
Objectives:
The goals of this work proposal are
two-fold: (i) building high-resolution multiscale panoramic mosaics from pan-tilt-zoom cameras,
(ii) detecting intrusions on the mosaics.
Detailed description:
As noted in the introduction,
automatic target detection is a required feature of modern video surveillance
systems. Target detection on large areas involves using many cameras.
Alternatively, as proposed in this work, one can use pan-tilt-zoom (PTZ)
cameras, i.e. mobile cameras with the elevation, azimuth and zoom degrees of freedom,
which allow simultaneously observing large (typically) static scenarios and
seeing in detail small portions where e.g. intrusion events were detected.
Most PTZ cameras closely resemble
central cameras, in the sense that despite changing their elevation and azimuth
angles, the imaging does not suffer (significant) parallax effects. This
property, combined with (almost) static scenarios, allows constructing simple
background-representations of the environment as the so termed video mosaics.
Video mosaics consist of enlarged
images of the scene obtained through the composition of individual (smaller
field of view) images. When using PTZ cameras, mosaics are built iteratively by
rotating the camera to cover the complete field of view.
This work-proposal encompasses
therefore two main steps:
(i) Build high-resolution multiscale
panoramic mosaics from PTZ cameras - In order to build mosaics from PTZ
cameras, it is necessary to calibrate the cameras' internal parameters using video
information (better accuracy than using the internal readings of the camera
pose [Sinha04]) and then registering the views acquired at different
orientations.
(ii) Detect intrusions on the mosaics - Detecting intrusions on the
scene involves detecting changing pixels, pixel-grouping to regions and
regions-filtering to remove non-relevant too small detections [Boult01]. The
mosaic (background model) has to be updated in order to deal with e.g. with
illumination or scene content changes.
References:
[Sinha04]
[Hall05] "Comparison of target
detection algorithms using adaptive background models", D. Hall, J. Nascimento, P. Ribeiro, E.
Andrade, P. Moreno, S. Pesnel, T. List, R. Emonet, R.B. Fisher, J. Santos Victor, J. Crowley, Int.
workshop on Performance evaluation of Tracking and Surveillance - Oct 2005
[Boult01] “Into the woods: Visual
surveillance of non-cooperative camouflaged targets in complex outdoor
settings,” T. Boult, R. Micheals,
X. Gao, and M. Eckmann,
Proc. IEEE, pp. 1382–1402, Oct. 2001.
[Nascimento06] "Performance
Evaluation of Object Detection Algorithms for Video Surveillance", Nascimento, J.C.; Marques, J.S.; IEEE T-Multimedia, V8.4,
2006 pp:761 - 774.
Requirements (Requisitos, e.g. média,disciplinas
concluídas):
Expected results (Resultado esperado):
At the end of the work, the students
will have enriched their experience in computer vision applied to video
surveillance setups. In particular are expected to:
- developed geometric modelling
methods for pan-tilt-zoom cameras
- algorithms
for intrusion detection and tracking.
Workplace (Local realização da dissertação):
ISR / IST
More MSc dissertation
proposals on Computer and Robot Vision in:
http://omni.isr.ist.utl.pt/~jag