|
With facial recognition, the overall performance
depends heavily on the quality of the input image, in particular
the image resolution of the facial area within the image, typically
measured in the number of pixels of inter eye distance (IED),
contrast or colour depth within the facial area, sharpness of
the objects within the image and pose of the facial image.
The system architecture design work showed that
standard CCTV would not be sufficient to give high enough resolution
for facial resolution algorithms to be used in the SUBITO scenarios
and thus the use of a Narrow Field Of View (NFoV) PTZ camera to
collect high resolution image was defined.
Image resolution is one of the most important
limiting factors with standard video-surveillance cameras. In
most cases, low resolution prevents object detection at a distance
exceeding 15 meters, and prevents face identification in nearly
all situations.
The work carried out in the SUBITO project has
shown that face recognition can be successfully achieved only
in cases of moderate poses (< 25° from head on) with inter-eye
distance larger than 20 pixels. Consequently, cameras dedicated
to face recognition should be located in a way that can capture
channelled person flows (e. g. stairs, passages, entrance and
exit points, etc.)
Low IED values normally result in a significant
degradation of the face recognition accuracy which would be
unacceptable for a SUBITO system. Thus the project developed a
concept that would improve recognition performance even in the
presence of such variations of IED.
The basic concept implemented was to incrementally
improve the facial representation of a person by adding in further
sample images of the same individual into the search image or
"probe" of the surveillance data.
Facial Detection Using Single Probe Image
For example, starting with a single image of a
person (probe) after an unattended baggage event was declared,
the system will extract faces from the video streams of the
remaining cameras and sort them according to decreasing similarity
of the person to be searched for, as illustrated in Figure 2 10.
In the case of high similarity values the system will (automatically
or with the support of an operator) add corresponding images to
the set of probe images. After this, the search may be repeated
or continued with the enlarged set of probe images leading to
possible further hits, Figure 2 11.
Facial Detection Using Multiple Probe Images
In this way the system can iteratively improve
the representation of a person covering more and more variations
of that individual and thus increasing the probability of recognition
at a different camera location and thus to track it across a
camera network.
Another key development to come from the work
carried out in the SUBITO project was the improvement of the
accuracy of the face recognition in the context of the scenario
where a person is tracked / recognised across a camera network.
In this scenario, the representation of a person differs
significantly from camera to camera due to different head poses,
varying illumination, camera characteristics, facial resolutions
etc.
|
