Information Technology Reference
In-Depth Information
3.2
Face Tracking
As we are examining video sequences rather than still images, we can exploit the
temporal relationships between frames to improve recognition accuracy. It is pos-
sible to track the movement of individual faces in a video stream using techniques
such as Boosted Particle Filtering, Mean Shift, or Kalman Filtering 6 .
If we can successfully track a moving face through a video sequence, we can take
each frame as being a sample of the same person. Extracted features can be tagged
with an identifier before sending to the back-end for recognition. Features with the
same identifier can be considered as multi-samples of the same individual.
Note: Multi-sampling was not implemented in the experimental system described
in this chapter.
3.3
Feature Extraction
One of the objectives of the face recognition system is invariance to changes in
pose. In previous approaches, this problem has been addressed by either creating
a 3D model [1] or by employing a pose estimation step [13, 32]. The creation of
a 3D model is problematic, as it requires a large set of training images for each
individual. This is impractical in a real-time surveillance application. Our system
follows a novel method, proposed in [33], where the feature vector is invariant to
pose, so a pose estimation step is not required.
[31] notes that the face detector can detect faces which are tilted up to about
15
±
45 out-of-plane (towards a profile view). In order to recognise faces
from any angle, our system constructs two separate models of each individual: a
frontal model and a profile model.
Each model (frontal and profile) defines a baseline ,a reference length and a
set of feature key points . The feature vector is defined by the spatial relationships
between these features (see section 4.1).
in plane and
±
3.3.1
Frontal Face Model
For the front face model, we first define two reference points — the centre points of
the irises. The front baseline is defined as a line connecting the reference points (see
figure 6. This length of this line is used as the reference length for the front face.
Next, we define a number of key points which denote the location of shape-based
features such as the nostrils and the tips of the ears. The feature vector is calculated
from a set of lengths and angles measured from the baseline to the feature key points.
The outline algorithm for finding the baseline and key feature points is as follows
(see figure 3):
6
Linear dynamic models are discussed in [7]. Non-linear models are also dis-
cussed in the unpublished chapter, “Tracking with non-linear dynamic mod-
els”, available at http://luthuli.cs.uiuc.edu/ daf/book/bookpages/
pdf/particles.pdf. See also [11].
Search WWH ::




Custom Search