The Problem:
Consider a crowded scene such as an airport. When
trying to identify people from distance using traditional
CCTV cameras, it will likely be unsuitable to use facial
recognition or other common biometrics such as gait,
ngerprint and iris recognition. It is therefore much more
suitable to build descriptors using features such as color
and texture. With these descriptors of individual people,
it is possible to track these individuals as they move
between non-overlapping cameras. Some researcher
presented a system based on producing a color silhouette
of a person, as an individuals appearance is dominated by
the clothes that they are wearing. From this, a color
histogram was used and compared with other histograms
produced from peoples silhouettes. Comparison was
achieved by nding out the histogram intersection, and
similar histogram intersections implied a match between
people in different images. However, color in an image
can be affected by many external factors, such as lighting
and the sensor in the camera; this can be resolved by
normalizing the image in order to provide consistency
between all cameras. The task of determining whether a
person passing through a network of camera has already
been observed is known as the people re-identication
problem, and is the focus of this research.
Current methods generally involve using humans in order
to monitor incoming surveillance video streams. the large
number of people being monitored would make this task
too large for human operators themselves to handle. In
addition, the accuracy of humans may be too low. This
research proposes a rank-style system, where the system
will produce a top n potential matches for an arbitrary
individual, and a human operator can pick from these
potential matches. This is known as a rank-n matching
scale, and is commonly used in this area. Efcient
methods to identify accurately are therefore an important
area of research.
DATA ACQUISITION
HUMAN
DETECTION
Motivation:
Aims:
Research challenges:
Appearance matching of the same person registered in
disjoint camera views is one of the most challenging
issues in every video analysis system. The underlying
challenge of the human re-identication problem arises
from signicant appearance changes, caused by variations
in viewing angle, illumination and object pose. Different
color responses, dierent camera viewpoints and
dierent camera parameters impede establishing
correspondence between parts of the human body, while
matching appearances.
Illumination changes: Day light intensity, shade,
changes in illumination color in indoor and outdoor
environments
Color changes: The same object acquired by dierent
cameras shows color dissimilarities
Pose changes: camera and viewpoint change as well as
the articulation of the human body lead to signicant
dierences in appearances of the same individual
observed from dierent cameras
Occlusions: self-occlusions (caused by body parts) and
occlusions caused by other people or objects
Scale changes: side of object depends on distance
between camera and object which may vary in
surveillance
Detections: to recognize individuals and detect human
silhouettes automatically
Discriminative features: to focus on distinctive cues
(features), providing within-group discriminations.
HUMAN
TRACKING
Research Impact:
There are several different ways of building a descriptor
of people, with examples being not only based on color,
but also on texture and soft biometrics. In addition, some
research has looked into how to identify people as groups
rather than individually However, no approach
particularly stands out above the rest. This research will
make it easier to track individuals by producing a rank-nstyle ranking. This will allow suspicious individuals to be
tracked in congested environments where this has
otherwise been challenging. Whilst most systems use
basic technologies such as color, which is easily inuenced
by illumination changes or background noise, our system
will isolate the background and take illumination changes
into account in order to attempt to prevent this. In
addition, the use of a skeleton model enables the system
to be aware of locations of particular parts of a persons
body. This aids not only in background subtraction, but
also in terms of tracking. Trajectory calculations could be
used to attempt to calculate a persons location even when
they are not in view of a camera.
Query to Database
SIGNATURE
COMPUTTATION
SIGNATURE
COMPUTTATION
SIGNATURE
DATABASE
Result
VIPeR
PRID
GRID
CUHK01
i-LIDs
ETHZ