STATISCAL-BASED SKIN CLASSIFIER FOR

OMNI-DIRECTIONAL IMAGES

Asaf Shupo, Bill Kapralos and Miguel Vargas Martin

University of Ontario Institute of Technology

2000 Simcoe Street North, Oshawa, Ontario, Canada. L1H 7K4

Keywords:

Skin detection, surveillance, omni-directional video sensing, maximum likelihood estimator.

Abstract:

This paper describes the development of a simple, video-based system capable of efﬁciently detecting human

skin in images captured with an omni-directional video sensor. The video sensor is used to provide a view of

the entire visual hemisphere thereby providing multiple dynamic views of a scene. Color models of both skin

and non-skin were constructed with images obtained with the omni-directional video sensor. Using a stochastic

weak estimator coupled with a linear classiﬁer, the system is capable of distinguishing omni-directional images

that contain human skin from those that do not. Results indicate that the system is able to accomplish this task

in a simple and computationally efﬁcient manner. The ability to obtain an image of the entire scene from a

single viewpoint using the omni-directional video sensor and determine whether the image contains human

skin (e.g., one or more humans) in a simple and efﬁcient manner is practical as a precursor for a number of

applications including teleconferencing, remote learning, and video surveillance, the application of interest in

this work.

1 INTRODUCTION

The area of human detection in the visual domain is

rather large, well investigated, and has many practi-

cal applications including surveillance (Boult et al.,

1998), video teleconferencing (Kapralos et al., 2003),

and face detection (Herpers et al., 1999). A good,

economical human detection/tracking system must be

able to locate humans quickly and reliably in the pres-

ence of noise and other objects in the environment.

It must run fast, efﬁciently (e.g., run in real-time),

and operate using inexpensive camera and computer

equipment (Bradski, 1998). Skin color is often pro-

posed as an economical and efﬁcient cue to detecting

humans in the visual domain. Color is the simplest at-

tribute in a set of pixels comprising the image (Jones

and Rehg, 1998) and does not require extensive com-

putational processing to compute. This allows a sys-

tem utilizing color cues to operate in real-time. In ad-

dition, the color of an object may be used as an iden-

tifying feature that is local to the object and largely

independent of the view and resolution. As a result,

the use of color information may be used to detect ob-

jects from differing viewpoints (Swain and Ballard,

1991). In general, color cues are invariant to par-

tial occlusion, rotation in depth, scale and resolution

changes (Raja et al., 1998). Furthermore, there are

various fast and simple color-based human detection

and tracking systems available (Chai and Ngan, 1999;

Chopra et al., 2006; Herpers et al., 1999; Jones and

Rehg, 1998; Kapralos et al., 2003; Shupo et al., 2006;

Stiefelhagen et al., 1999; Hans et al., 1999; Yang and

Waibel, 1996).

Although skin detection is itself a simple and ef-

ﬁcient process, many of the existing systems employ

traditional cameras with a limited ﬁeld of view. Us-

ing such cameras, in order to capture a view of the

entire visual hemisphere, multiple stationary cameras

may be used or a single camera may be panned to

different directions. Furthermore, in various appli-

cations (such as video teleconferencing) the subject

(person) may physically move into the view of a sta-

tionary camera. Both approaches can greatly increase

the computational and time requirements potentially

making these approaches impractical for real-time op-

erations. Rather than having a user move into the

camera’s ﬁeld of view, having multiple cameras or

focusing the camera in different directions, an omni-

Shupo A., Kapralos B. and Vargas Martin M. (2007).

STATISCAL-BASED SKIN CLASSIFIER FOR OMNI-DIRECTIONAL IMAGES.

In Proceedings of the Second International Conference on Computer Vision Theory and Applications, pages 11-17

DOI: 10.5220/0002060800110017

 SciTePress

directional video sensor (Cyclovision’s ParaCamera

system (Nayar, 1997; Baker and Nayar, 1999)) can

be utilized instead. The ParaCamera captures a 360

◦

(hemispherical) view from a single viewpoint.

In this paper we describe an approach that is used

to classify ParaCamera images as either i) containing

skin or ii) not containing skin. Color models of both

skin and non-skin were constructed with images ob-

tained with the ParaCamera. Using a stochastic esti-

mator coupled with a linear classiﬁer, results suggest

the system is capable of distinguishing images that

contain human skin from images that do not. This

work is part of an ongoing research project investigat-

ing the fundamental issues related to the development

of a video-based surveillance and monitoring system

capable of locating humans within a scene. Humans

within a scene may represent potential intruders and

allows them to be automatically detected in an efﬁ-

cient manner can allow further, more complex actions

to be taken. Such actions may include focusing a high

resolution pan-tilt camera on to the potential intruder,

and alerting a human operator.

The remainder of the paper is organized as fol-

lows. Section 2 provides greater details regarding the

proposed system. In particular, further details regard-

ing the ParaCamera and a detailed description of the

statistical skin detection method is provided. Results

of several experiments conducted to provide an indi-

cation of the effectiveness of method are provided in

Section 3. Finally, concluding remarks and plans for

future research are presented in Section 4.

2 THE APPROACH

2.1 ParaCamera Omni-Directional

Camera System

Cyclovision’s ParaCamera omni-directional camera

consists of a high precision paraboloidal mirror and

a combination of special purpose lenses (see Figure

1). By aiming a suitably equipped camera at the face

of the paraboloidal mirror, the optics assembly per-

mits the ParaCamera to capture a 360

hemispherical

view from a single viewpoint. Once the hemispherical

view has been obtained, it may be easily un-warped

(Peri and Nayar, 1997) producing a panoramic view.

From this panoramic view, a perspective view of any

size corresponding to portions of the scene can be eas-

ily extracted. An example of both hemispherical and

panoramic images are illustrated in Figure 2. Ex-

amination of the images shown in Figure 2 illustrates

the distortion and varying resolution inherent with im-

Paraboloidal

mirror

Paracamera lens

assemebly

Digital

camera

Figure 1: Cyclovision’s ParaCamera omni-directional video

sensor. The camera is aimed at the face of a paraboloidal

mirror allowing it to capture a view of the entire visual

hemisphere from a single viewpoint.

directional video sensor (Cyclovision’s ParaCamera

system (Nayar, 1997; Baker and Nayar, 1999)) can

be utilized instead. The ParaCamera captures a 360

◦

(hemispherical) view from a single viewpoint.

In this paper we describe an approach that is used

to classify ParaCamera images as either i) containing

skin or ii) not containing skin. Color models of both

skin and non-skin were constructed with images ob-

tained with the ParaCamera. Using a stochastic esti-

mator coupled with a linear classiﬁer, results suggest

the system is capable of distinguishing images that

contain human skin from images that do not. This

work is part of an ongoing research project investigat-

ing the fundamental issues related to the development

of a video-based surveillance and monitoring system

capable of locating humans within a scene. Humans

within a scene may represent potential intruders and

allows them to be automatically detected in an efﬁ-

cient manner can allow further, more complex actions

to be taken. Such actions may include focusing a high

resolution pan-tilt camera on to the potential intruder,

and alerting a human operator.

The remainder of the paper is organized as fol-

lows. Section 2 provides greater details regarding the

proposed system. In particular, further details regard-

ing the ParaCamera and a detailed description of the

statistical skin detection method is provided. Results

of several experiments conducted to provide an indi-

cation of the effectiveness of method are provided in

Section 3. Finally, concluding remarks and plans for

future research are presented in Section 4.

2 THE APPROACH

2.1 ParaCamera Omni-Directional

Camera System

Cyclovision’s ParaCamera omni-directional camera

consists of a high precision paraboloidal mirror and

a combination of special purpose lenses (see Figure

1). By aiming a suitably equipped camera at the face

of the paraboloidal mirror, the optics assembly per-

mits the ParaCamera to capture a 360

hemispherical

view from a single viewpoint. Once the hemispherical

view has been obtained, it may be easily un-warped

(Peri and Nayar, 1997) producing a panoramic view.

From this panoramic view, a perspective view of any

size corresponding to portions of the scene can be eas-

ily extracted. An example of both hemispherical and

panoramic images are illustrated in Figure 2. Ex-

amination of the images shown in Figure 2 illustrates

the distortion and varying resolution inherent with im-

ages obtained using a ParaCamera. In the hemispher-

ical view (see Figure 2(a)), the resolution decreases

Paraboloidal

mirror

Paracamera lens

assemebly

Digital

camera

Figure 1: Cyclovision’s ParaCamera omni-directional video

sensor. The camera is aimed at the face of a paraboloidal

mirror allowing it to capture a view of the entire visual

hemisphere from a single viewpoint.

(a) Hemispherical view.

(b) Panoramic view

Figure 2: Sample ParaCamera images. (a) Hemispherical

view. (b) A simple procedure allows the hemispherical view

to be transformed into a panoramic view.

as we move away from the center of the image. Fur-

ther examination also reveals that given the low reso-

lution inherent in these images, ﬁne details (e.g., fa-

cial features) cannot generally be detected. This can

be a limitation with many of the existing computer vi-

sion/image processing algorithms currently available

(e.g., face detection by locating both eyes). However,

the low resolution inherent with ParaCamera images

does not pose a problem for this application. As will

be described below, in this work the goal is to em-

(a) Hemispherical view.

(b) Panoramic view

Figure 2: Sample ParaCamera images. (a) Hemispherical

view. (b) A simple procedure allows the hemispherical view

to be transformed into a panoramic view.

ages obtained using a ParaCamera. In the hemispher-

ical view (see Figure 2(a)), the resolution decreases

as we move away from the center of the image. Fur-

ther examination also reveals that given the low reso-

lution inherent in these images, ﬁne details (e.g., fa-

cial features) cannot generally be detected. This can

be a limitation with many of the existing computer vi-

sion/image processing algorithms currently available

(e.g., face detection by locating both eyes). However,

the low resolution inherent with ParaCamera images

does not pose a problem for this application. As will

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

be described below, in this work the goal is to em-

ploy a ParaCamera to obtain a fast overview of the

scene while ﬂagging potential areas of human activ-

ity. Other higher resolution imaging sensors can then

be focused on these potential areas of interest in order

to provide further information.

The ability of an omni-directional video sensor

such as the ParaCamera to capture an image of the

entire visual hemisphere makes it very attractive for

a variety of applications. Such applications include

the capture the simultaneous video of each partici-

pant in a small group meeting (Stiefelhagen et al.,

1999; Yong et al., 2001), surveillance (Boult et al.,

1998; Gutchess et al., 2000), autonomous robot nav-

igation (Zheng and Tsuji, 1992), virtual reality (Ya-

sushi, 1999), telepresence (Yasushi, 1999), video-

teleconferencing (Kapralos et al., 2003) and pipe in-

spection (Basu and Southwell, 1995).

2.2 Skin Classiﬁcation

Skin detection is performed using an estimator for

feature extraction coupled with a linear classiﬁer. The

estimators considered in this work are the stochas-

tic learning weak estimator (SLWE) (Oommen and

Rueda, 2006) and the maximum likelihood estimator

(MLE) (Duda et al., 2000). The SLWE is considered

to be more accurate in dealing with non-stationary

data (e.g., a sequence of images capturing the mo-

tion of a subject), which is a relevant consideration

when considering video surveillance. The classiﬁca-

tion process consists of two phases: i) the training

phase and ii) the testing phase. In the training phase,

two feature vectors are used to classify the skin and

non-skin images (e.g., images that contain human skin

regions and images that do not respectively). By ex-

tracting statistical properties from the labeled images

(i.e., those that have been sorted into the skin and

non-skin vectors), we are able to conduct the exper-

iments using this as a basis of comparison. The fea-

tures that were extracted from the training phase with

any needed adjustments were input into the classiﬁer

that we used in the validation phase of the SLWE. The

experiments were then repeated for the MLE (the es-

timators are discussed in the following sections). The

training phase, as pointed out above, aims ﬁrst to ex-

tract the statistical features of the images correspond-

ing to all images in the training dataset, producing one

feature vector for each class. The procedure shown in

Algorithm 1 produces these two feature vectors when

it is run for each dataset (e.g., skin and non-skin).

The algorithm is used separately for the skin and

non-skin training datasets. The output of the algo-

rithm is a feature vector, an array V

◦

or V

, one for the

Algorithm 1 Procedure for determining the fea-

ture vectors.

1: Initialize an array B of counters to zero

2: For each image I of the training dataset of class j:

3: For each 8-bit byte b

of I:

4: Increment B[b

] by 1

5: Initialize an array V

of probabilities to zero.

6: For k = 0 to 255

7: Set V

[k] = B[b

] / total number of 8-bit bytes of

the set of images

skin and non-skin dataset, respectively. The appro-

priate statistical characteristics are initially extracted

and the feature vectors V

◦

and V

are formed. The

next step is to use an estimator to extract the features

of the image to be classiﬁed, namely a vector V

. The

classiﬁcation rule consists of assigning an unlabeled

package to the class, skin or non-skin, that minimizes

the distance between V

and the trained arrays V

◦

. Two metrics have been used for this purpose with

both the ML and SLWE algorithms (described later

in greater detail): i) the Euclidean distance, and ii)

the weighted Euclidean distance. Both metrics are

used to calculate the distance between the two labeled

vectors, and once the distance has been determined, a

classiﬁcation is made as to whether an image contains

skin or non-skin (e.g., the image is classiﬁed as being

either “skin” or “non-skin”). Before using a classiﬁ-

cation metric, statistical characteristics of the datasets

must be extracted. In the following sections, the MLE

and SLWE used to extract such features are described

in greater detail. For this purpose, we obtain the fre-

quency of occurrence for each of the symbols (from 0

to 255) for a given image that has to be classiﬁed.

2.2.1 The Maximum Likelihood Estimator

The maximum likelihood estimator (MLE) is a tradi-

tional technique that aims to maximize the likelihood

that a given sample generates using a speciﬁc prob-

abilistic model, either parametric or non-parametric.

We assume that we are dealing with a multinomial

random variable with 256 possible realizations (one

symbol for each 8-bit ASCII value). It has been

shown that the likelihood is maximized when the esti-

mate for each symbol is given by the frequency coun-

ters divided by the total number of bytes in the image

(Duda et al., 2000) (see Algorithm 2).

The algorithm produces an array V

that contains

the estimates for each (8-bit) byte in the testing image

H. That vector V

is then input to the classiﬁcation

rule, which decides on the class based on a distance

function and the trained feature vectors.

STATISCAL-BASED SKIN CLASSIFIER FOR OMNI-DIRECTIONAL IMAGES

Algorithm 2 The maximum likelihood estimator.

1: For each image H captured by the ParaCamera:

2: Initialize an array C of counters to zero

3: For each 8-bit byte b

of H:

4: Increment C[b

] by 1

5: Initialize an array V

of probabilities to zero

6: For k = 0 to 255

7: Set V

[k] = C[b

] / total number of 8-bit bytes of

this image.

Algorithm 3 The stochastic learning weak esti-

mator.

1: For each image H captured by a ParaCamera:

2: Initialize each entry of the feature

vector V

to 1/256

3: For each 8-bit byte b

of H:

4: For k = 0 to 255

5: If i 6= bi then

6: V

[k] = λ ×V

[k]

7: Else

8: V

] = V

] + (1 − λ)

∑

k6=i

[k]

2.2.2 The Stochastic Learning Weak Estimator

Estimators like the one described by the MLE algo-

rithm outlined in the previous section suffer from a

lack of ability to capture quick changes in the dis-

tribution of the source data (e.g., dealing with non-

stationary data, that is, data from different types of

scenarios). Oommen and Rueda (2006) proposed

a stochastic learning weak estimator (SLWE). The

SLWE combined with a linear classiﬁer has been suc-

cessfully used to deal with problems that involve non-

stationary data and has been effectively used to clas-

sify television news into business and sports news

(Oommen and Rueda, 2006). In this work, each im-

age to be classiﬁed is read from the testing dataset,

and is used to feed the classiﬁcation rule by means

of extracting statistical features into a feature vector.

The source alphabet contains n symbols (n = 256),

which represent the possible realizations of a multino-

mial random variable, and whose estimates are to be

updated by using the SLWE rules. While this rule re-

quires a ”learning” parameter (λ), it has been found

that a good value for multinomial scenarios should

be close to 1 (e.g., λ = 0.999 (Oommen and Rueda,

2005)) (see Algorithm 3).

The classiﬁcation rule is validated using labeled

images and adjustments are made if necessary. Note

that in the actual classiﬁcation process the label of

each image is not known. To classify the complete im-

age, several distance metrics are employed. Greater

details regarding these metrics are provided in the fol-

lowing section.

2.2.3 Distance Metrics

The choice of a distance function (also referred to

as “metric”) is not a trivial task. Often, different

components of the feature vectors may have different

weights in classiﬁcation of an arbitrary image. Some

entries of the feature vector may be more important

than other entries, or some entries may have more

noise than other entries. Therefore, the choice of a

metric plays an important role in the performance of

the algorithm.

Euclidean Distance In this metric, it is assumed

that all entries in the feature vector have equal weight.

The Euclidean distance d(V,V

) between two feature

vectors V and V

is deﬁned as

d(V,V

) =

255

∑

i=0

(V [i] −V

[i])

(1)

Weighted Euclidean Distance This metric is also

known as the Mahalanobis distance when the covari-

ance matrix is considered as a diagonal matrix. It

is assumed that different entries in the feature vec-

tor have different importance in the classiﬁcation of

images. It is also assumed that an entry in the feature

vector is of less importance than another entry if its

variance is greater than the variance of another entry.

The weighting factor w is deﬁned as w = 1/σ

, and

the weighted Euclidean distance is given as

d(V,V

) =

255

∑

i=0

(V [i] −V

[i])

(2)

A discussion regarding some of the issues related to

the weighting factor w are presented in (Shupo et al.,

2006).

3 EXPERIMENTAL RESULTS

In this section, the results of two experiments are pre-

sented. For both experiments two datasets of im-

ages obtained with the ParaCamera were used. The

ﬁrst dataset contained 43 images that contained one

or more humans in the scene and therefore skin re-

gions (e.g., “skin images”). From these images, 24

were used in the learning phase and 19 were used

in the testing phase. The second dataset contained

47 images without any humans present in the scene

and therefore no skin regions within the images (e.g.,

“non-skin images”). From these images, 24 of them

were used in the learning phase and 23 were used in

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

the testing phase. The images in both datasets were

of type JPEG. JPEG images are compressed images

where each “byte segment” contains the information

needed to reconstruct the original image. For exam-

ple, a byte or group of bytes may represent the encod-

ing of some coefﬁcient that results from the discrete

cosine transform, or other transformation. In some

cases, however, the encodings result in variable length

codes (i.e., of length not necessarily multiple of 8)

which are spread out in a number of bytes. Sample

images from both the skin and non-skin datasets are

provided in Figure 3. Both the ML and SLWE employ

the use of a distance measure and appropriate distance

thresholdin order to classify an image as either “skin”

or “non-skin”. In general, different distance metrics

and different threshold settings may lead to different

results leading to variations in the number of false

positives and false negatives. In order to draw a mean-

ingful comparison between the results of the ML and

SLWE algorithms using the Euclidean and weighted

Euclidean distance metrics, results are presented as

a comparison between the probability of false posi-

tives (e.g., the probability of incorrectly classifying

an image as ”skin” when in fact it does not contain

any skin) and the probability of false negatives (e.g.,

the probability of incorrectly classifying an image as

”non-skin” when in fact it does contain skin). In other

words, setting the algorithmic parameters such that

the algorithm results in a particular number of false

negatives implies the algorithm will also result in the

corresponding number of false positives. The result-

ing graphs illustrate the probability of false negatives

vs. the probability of false positives. A summary of

the results for the ML and SLWE algorithms using

the weighted Euclidean distance metric are illustrated

in Figure 4(a),(b). The resulting graphs illustrate the

probability of false negatives vs. the probability of

false positives.

Since there is a trade-off between the probability

of false positives and false negatives with each algo-

rithm, one algorithm may outperform the other (e.g.,

provide “better” results, where “better” is deﬁned as

minimizing either the number of false positives, the

number of false negatives or having an equal number

of false positives and false negatives). When consid-

ering the weighted Euclidean distance, the ML algo-

rithm with a 0% false positive rate and a 12% false

negative rate, outperforms the SLWE when a mini-

mal false positive rate is desired. With respect to the

Euclidean distance, the SLWE with a false positive

rate of 17% and false negative of 12% provides su-

perior results over the ML algorithm when an equal

number of false positive and false negatives is desired.

A summary of the results for the ML and SLWE algo-

(a) “Skin” image.

(b) “Non-skin” image.

Figure 3: Sample images form the “skin” and “non-skin”

image datasets. (a)“Skin image” (e.g., the image contains

regions of human skin) and (b) “non-skin image” (the image

does not contain any regions of human skin).

rithms using a weighted Euclidean distance metric are

illustrated in Figure 5(a),(b). The resulting graphs il-

lustrate the probability of false negatives vs. the prob-

ability of false positives. Since there is a trade-off be-

tween the probability of false positives and false nega-

tives with each algorithm, one algorithm may outper-

form the other (provide “better” results where “bet-

ter” is with respect to minimizing the number of both

false positives and false negatives). With respect to

the Euclidean distance metric, the SLWE with a false

positive rate of 17% and false negative of 12% pro-

STATISCAL-BASED SKIN CLASSIFIER FOR OMNI-DIRECTIONAL IMAGES

Decision Strategy For ML Algorithm in Euclidean Distance

0.2

0.4

0.6

0.8

0 0.1 0.2 0.3 0.4

Probability (False Negative)

Probability (False Positive)

(a) ML with the Euclidean distance metric.

Decision Strategy For SLWE Algorithm in Euclidean Distance

0.2

0.4

0.6

0.8

0 0.1 0.2 0.3 0.4

Probability (False Negative)

Probability (False Positive)

(b) SLWE with the Euclidean distance metric.

Figure 4: Decision strategies. Probability of false posi-

tives (e.g., the probability of incorrectly classifying an im-

age as “skin” when in fact it does not contain any skin) vs.

the probability of false negatives (e.g., probability of incor-

rectly classifying an image as “non-skin” when in fact it

does contain skin) for the ML and SLWE algorithms em-

ploying the Euclidean distance metric. (a) ML algorithm

and (b) SLWE algorithm.

vides superior results over the ML algorithm. When

considering the weighted Euclidean distance metric,

the ML algorithm with a weighted Euclidean distance

with a 0% false positive rate and a 12% false nega-

tive rate provides superior results over the SLWE al-

gorithm.

4 CONCLUSION

In this paper a description of a color-based human

skin detector was presented. The system is capable of

accurately and efﬁciently classifying images obtained

with an omni-directional video sensor as either con-

taining skin or not containing skin. The system is part

of a larger ongoing project whose goal is to develop an

autonomous surveillance system that can monitor an

area for human activity. The work described here pre-

sented one aspect of this system. In particular, the de-

velopment of a video sensor to obtain a fast overview

of the scene and identify potential areas of human ac-

tivity. Various other techniques can then be focused to

these potential areas of interest in order to obtain fur-

Decision Strategy For ML Algorithm in Weighted Euclidean

Distance

0.2

0.4

0.6

0.8

0 0.1 0.2 0.3 0.4

Probability (False Negative)

Probability (False Positive)

(a) ML with the weighted Euclidean distance metric.

Decision Strategy For SLWE Algorithm in Weighted Euclidean

Distance

0.2

0.4

0.6

0.8

0.00 0.10 0.20 0.30 0.40

Probability (False Negative)

Probability (False Positive)

(b) SLWE with the weighted Euclidean distance metric.

Figure 5: Decision strategies. Probability of false posi-

tives (e.g., the probability of incorrectly classifying an im-

age as “skin” when in fact it does not contain any skin) vs.

the probability of false negatives (e.g., probability of incor-

rectly classifying an image as “non-skin” when in fact it

does contain skin) for the ML and SLWE algorithms em-

ploying the weighted Euclidean distance metric. (a) ML

algorithm and (b) SLWE algorithm.

ther details. Two algorithms were presented: the max-

imum likelihood estimator and the stochastic learning

weak estimator and for each algorithm, two distance

metrics (Euclidean and weighted Euclidean) were ex-

perimented with.

During a training phase, ParaCamera images are

classiﬁed as either skin (contain regions of skin) or

non-skin (do not contain any regions of skin). The

features that are extracted from the training phase

with any needed adjustments are input into the clas-

siﬁer that was later used in the validation phase to

classify incoming images as skin or non-skin. Re-

sults were presented that compared both algorithms

and both distance metrics based on the number of

false positives and false negatives. The ML algorithm

provided superior results using the Euclidean distance

metric while the SLWE provided superior results us-

ing a weighted Euclidean distance metric. Results

also indicate that the system is capable of accurately

classifying ParaCamera images as either skin or non-

skin in a very efﬁcient manner even when considering

the poor resolution inherent with ParaCamera images.

Future work will include greater, more extensive

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

experimentation using much larger training datasets

and a larger number of test images obtained that have

been obtained under a variety of scenarios and light-

ing conditions. Future work will also examine the

processing of skin classiﬁed images in order to ob-

tain greater details regarding the photographed scene.

Processing can include locating and grouping skin re-

gions within the image and inferring their position in

the real world. This information is of importance to

any surveillance application.

ACKNOWLEDGEMENTS

Asaf Shupo and Miguel Vargas Martin gratefully ac-

knowledge the ﬁnancial support of the Natural Sci-

ences and Engineering Research Council of Canada

(NSERC).

REFERENCES

Baker, S. and Nayar, S. (1999). A theory of single viewpoint

catadioptric image formation. International Journal

of Computer Vision, 35(2):1–22.

Basu, A. and Southwell, D. (1995). Omni-directional sen-

sors for pipe inspection. In IEEE Transactions on Sys-

tems Man and Cybernetics, volume 25, pages 3107–

3112.

Boult, T., Michaels, R., Gao, P., Lewis, C.,

Yin, W., and Erkan, A. (1998). Frame

rate omni-directional surveillance and track-

ing of camouﬂaged and occluded targets.

http://www.eecs.lehigh.edu/ tboult/TRACK/LOTS.html.

Bradski, G. (1998). Computer vision face tracking for use

in a perceptual user interface. Technical report, Intel

Corp., Santa Clara, CA USA.

Chai, D. and Ngan, K. (1999). Face segmentation us-

ing skin-color map in videophone applications. IEEE

Transactions on Circuits Systems and Video Technol-

ogy, 9(4):551–564.

Chopra, M., Vargas Martin, M., Rueda, L., and Hung, P.

(2006). A source address reputation system to com-

bating child pornography at the network level. In

Proceedings of the IADIS International Conference on

Applied Computing, San Sebastian, Spain.

Duda, R. O., Hart, P. E., and Stork, D. E. (2000). Pat-

tern Classiﬁcation. Wiley Interscience, Hoboken, NJ.

USA, second edition.

Gutchess, D., Jain, A., and Cheng, S. (2000). Automatic

surveillance using omni-directional and active cam-

eras. In Proceedings of the 2000 Asian Conference

on Computer Vision.

Hans, S., Anderson, H., and Granum, E. (1999). Skin color

detection under changing lighting conditions. In Pro-

ceedings of the 7th Symposium on Intelligent Robotics

Systems,, pages 187–195, Columbia, Portugal.

Herpers, R., Derpanis, K., Topalovic, D., and Tsotsos, J.

(1999). Detection and tracking of faces in real envi-

ronments. In Proceedings of the International Work-

shop on Recognition, Analysis and Tracking of Faces

in Real-Time Systems, pages 96–104, Korfu, Greece.

Jones, M. and Rehg, J. (1998). Statistical color models

with applications to skin detection. Technical Report

CRL 98/11, Compaq Computer Corp., Cambridge,

MA USA.

Kapralos, B., Jenkin, M., and Milios, E. (2003). Audio-

visual localization of multiple speakers in a video tele-

conferencing setting. Journal of Imaging Science and

Technology, 13(1):95–105.

Nayar, S. (1997). Omnidirectional video camera. In Pro-

ceedings of the DARPA Image Understanding Work-

shop, pages 235–241, New Orleans, LA.

Oommen, B. J. and Rueda, L. (2005). On utilizing stochas-

tic learning weak estimators for training and classi-

ﬁcation of patterns with non-stationary distributions.

In Proc. 28th German Conf. on AI, pages 107–120,

Koblenz, Germany.

Oommen, B. J. and Rueda, L. (2006). On utilizing stochas-

tic learning weak estimators for training and classi-

ﬁcation of patterns with non-stationary distributions.

Journal of Pattern Recognition, 39:328–341.

Peri, V. and Nayar, S. (1997). Generation of perspective and

panoramic video from omnidirectional video. In Pro-

ceedings of the DARPA Image Understanding Work-

shop, pages 243–245, New Orleans, LA USA.

Raja, Y., McKenna, J., and Gong, S. (1998). Segmenta-

tion and tracking using color mixture models. In Pro-

ceedings of the Third Asian Conference of Computer

Vision.

Shupo, A., Vargas Martin, M., Rueda, L., Bulkan, A., Chen,

Y., and Hung, P. (2006). Toward efﬁcient detec-

tion of child pornography in the network infrastruc-

ture. IADIS International Journal on Computer Sci-

ence and Information Systems, 1(2):15–31.

Stiefelhagen, R., Yang, J., and Waibel, A. (1999). Mod-

eling focus of attention for meeting indexing. In Pro-

ceedings of the ACM Multemedia ’99, pages 3–10, Or-

lando, FL USA.

Swain, M. and Ballard, D. (1991). Color indexing. Interna-

tional Journal of Computer Vision, 7:11–32.

Yang, J. and Waibel, A. (1996). A real time face tracker.

In Proceedings of the 3rd IEEE Workshop on Applica-

tions of Computer Vision, Sarasota, FL, USA.

Yasushi, Y. (1999). Omni-directional sensing and its appli-

cations. IEEE Transactions on Information and Sys-

tems, E82-3.

Yong, R., Gupta, A., and Cadiz, J. (2001). Viewing meet-

ings captured by an omni-directional camera. In ACM

Transactions on Computer-Human Interaction.

Zheng, J. and Tsuji, S. (1992). Panoramic representation

for route recognition by a mobile robot. International

Journal of Computer Vision, 9(1):55–76.

STATISCAL-BASED SKIN CLASSIFIER FOR OMNI-DIRECTIONAL IMAGES