January 10, 2010 / Vol. 8, No. 1 / CHINESE OPTICS LETTERS 59
Eye location under different eye poses, scales, and
illuminations
Jinghe Yuan (µµµÚÚÚ)
Key Laboratory of Molecular Nanostructure and Nanotechnology, Institute of Chemistry,
Chinese Academy of Sciences, Beijing 100190, China
E-mail: jacobyuan@yahoo.com.cn
Received March 10, 2009
Robust non-intrusive eye location plays an important role in vision-based man-machine interaction. A
mo dified Hausdorff distance based measure to localize the eyes is proposed, which could tolerate various
changes in eye pose, shap e, and scale. To eliminate the effects of the illumination variations, an 8-
neighb our-based transformation of the gray images is proposed. The transformed image is less sensitive
to illumination changes while preserves the appearance information of eyes. All the localized candidates
of eyes are identified by back-propagation neural networks. Exp eriments demonstrate that the robust
metho d for eye location is able to localize eyes with different eye sizes, shapes, and poses under different
illuminations.
OCIS co des: 150.1135, 100.2000, 100.3008.
doi: 10.3788/COL20100801.0059.
Robust non-intrusive eye location plays an imp ortant role
in vision-based man-machine interaction including auto-
motive applications, such as driver inspection, face recog-
nition, etc. In the past years, many works were addressed
on this area. There are two major approaches for auto-
matic eye detection. The first approach, the holistic one,
conceptually relates to template matching, and attempts
to locate the eye using global representations. Character-
istic of this approach belongs to connectionist methods
such as principal component analysis (PCA) using eigen-
representations
[1]
. Although location by matching raw
images has been successful under limited circumstances,
it suffers from the usual shortcomings of straightforward
correlation-based approaches, such as sensitivity to eye
orientation, size, variable lighting conditions, noise, etc.
The second approach for eye detection extracts and mea-
sures local facial features, while standard pattern recog-
nition techniques are then employed for locating the eyes
using these measurements. Yuille et al. described a com-
plex but generic strategy
[2]
. The characteristic of this
approach is the concept of deformable templates. Lam et
al. extended Yuille’s method to extract eye features by
using corner locations inside the eye windows which are
obtained by means of average anthropometric measures
after the head boundary is located
[3]
. Deformable tem-
plate is an interesting concept, but it is difficult in terms
of learning and implementation to use them.
Hausdorff distance was originally defined as a dissimi-
larity measure on data sets. It later got wide acceptance
in image comparison. Huttenlocher et al. proposed a
partial Hausdorff distance (PHD) method for object de-
tection and recognition
[4]
. This method gets distance
measure between the most closely matching portions of
the images being compared which in turn reduces the
effect of occlusion in object matching. Guo et al. pro-
posed spatially weighted Hausdorff distance (SWHD) as
an improvement to conventional Hausdorff distance be-
tween edge images
[5]
.
All the above-mentioned methods for feature recogni-
tion use edge images for finding Hausdorff distance or
its variants. However, the appearance is more impor-
tant than the edge maps. The intensity distribution of
pixels captures this appearance information. However,
direct comparison of gray images is unsuitable because
the performance will be affected by illumination vari-
ations. To overcome this shortcoming, infrared imag-
ing technique
[6]
, method of single training image per
person
[7]
, and local binary patterns (LPBs)
[8]
were pro-
posed.
In this letter, by using the average instead of the maxi-
mum in the directed Hausdorff distance, a modified Haus-
dorff distance (MHD) based measure
[9]
for comparing
the appearance of eyes is proposed, which is able to tol-
erate changes in eye shape, pose, and size. To elimi-
nate the effects of the illumination variations, we pro-
pose an 8-neighbour-based transformation of the gray
images. The transformed eye image is less sensitive to
illumination changes while preserves the appearance in-
formation of eyes. All the located eyes are identified by
back-propagation neural network (BPNN) identifier.
Primarily, the Hausdorff distance is defined as a dis-
tance measure between two data sets. This distance
gives a measure of dissimilarity between these two sets.
The conventional Hausdorff distance between the two sets
A = {a
1
, a
2
,· · ·, a
m
} and B = {b
1
, b
2
,· · ·, b
n
} is given by
H(A, B) = max[h(A, B), h(B, A)], (1)
where
h (A, B) = max
a∈A
min
b∈B
ka − bk (2)
and
h (B, A) = max
b∈B
min
a∈A
kb − ak (3)
are the directed Hausdorff distances from A to B and
from B to A, respectively, and k·k is the norm of a vector.
H (A, B) takes the maximum of the directed distances
from A to B and from B to A. When the Hausdorff
distance is measured b etween two images, the data sets
1671-7694/2010/010059-04
c
° 2010 Chinese Optics Letters