Information Technology Reference
In-Depth Information
simple description methods can be sufficient, provided that camera colour calibration
techniques are used [ 27 ].
The actual object re-identification task involves implementation of the classifica-
tion scheme that allows determining which object candidate in a destination camera
is the one observed in a source camera. This task may be based on a simple, distance-
based measure in order to compare visual feature vectors of objects' images with
each other, or it may employ intelligent decision systems techniques, such as Neural
Networks or Decision Trees.
Surveillance systems can contain dozens or even hundreds of cameras. Searching
for the tracked object in each camera of such systems would be inefficient and could
lead to many errors. In order to solve this problem, information about the system
topology should be included in the process [ 18 ]. This way, if an object leaves the field
of view of a certain origin camera, its representation is required to be searched only
in the cameras with a proper spatio-temporal relation from the origin. For further
improvement of the object re-identification accuracy, time windows representing
expected transition times between cameras can be utilized. Their purpose is to define
additionally the most probable periods within which the objects are sought after.
Such a mechanism is the most beneficial in case when neighbouring cameras are
placed at a considerable distance between each other.
The third hint in the re-identification process is a behaviour model that utilizes
information on frequency of particular transitions between cameras. Moreover, the
complete paths (routes) of objects can be analysed in order to obtain a behaviour
model. Such a model contains a statistical description of objects behaviour, so it can
be used to predict future movement of the object or allow reconstructing its route in
the past. In literature, many methods of building a behaviour model are described.
Kettnaker et al. present utilization of Markov models [ 29 ]. The number of future steps
which can be predicted is related to the order of Markov model. Another method
presumes usage of particle filters based on previously collected statistical data about
routes of objects [ 30 ]. Behaviour model facilitating object re-identification might be
also based on the idea of Pawlak's flowgraphs [ 14 ]. The latter two algorithms use
large amount of statistical data as an input.
To sum up, the multi-camera object tracking process combines many types of
meta-data obtained from the consecutive steps of video analysis. Depending on the
changing conditions of video acquisition (different illumination types, number of
objects tracked by video surveillance system, amount of statistical data used for build-
ing behaviour model, modifications in the camera network topology etc.), importance
of a particular type of hints can also be modified.
12.4 Visual Object Descriptors
A large miscellany of visual object descriptors for image classification are available
[ 2 , 3 , 24 , 35 ]. However, only some of them can be utilized for the problem considered
here [ 5 , 32 , 44 ]. As it was stated in the previous Section, object description for the
purpose of re-identification in a multi-camera surveillance system is expected to be
 
Search WWH ::




Custom Search