International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 1

Counting People Using

Blobs and Contours
Shafraz Subdurally, University of Mauritius, Rduit, Mauritius
Devin Dya, University of Mauritius, Rduit, Mauritius
Sameerchand Pudaruth, University of Mauritius, Rduit, Mauritius

ABSTRACT
Counting the number of people in public locations has become imperative in surveillance applications for
the good management of public space. The automatic counting of people can indeed help carry out the above
tasks better and faster. In this paper, the authors propose two systems for counting people from images. Their
proposed methods are based on the observation that heads are significantly more visible than any other features
and are thus more easily distinguishable. The proposed systems use blobs and contour detection respectively
to count the number of people. The results obtained from each system are very reliable. The average head
detection rate of the systems is 82 and 84 percent respectively.

Keywords: Blob, Contour Detection, Crowd, Observation, Public Space

1. INTRODUCTION is a need to enlarge the location area for greater

People have always attended their preferred
political partys meeting, supported their fa-
vourite team at a sports event, demonstrated a
particular idea during a rally or simply doing
some shopping during the new years eve.
Manual counting of people in such places can
often be difficult since they are not arranged
in a specific order and they may be in constant
motion. The aim of counting people is numer-
ous and a few conventional ways of counting
people include using laser beams, infra-red
sensors, thermal sensors (Lefloch, 2007), etc.
Counting people in a particular location can
help in decision making such as to see if there
pedestrians safety (Anthony C. Davies, 1995).
Events held on streets like concerts and public
meetings make counting people more difficult.
People counting systems have evolved in recent
years from manual head counts to computer
vision neural learning but still most recent
systems do not perform efficiently where there
is a large number of people with occlusions
(Lim Eng Aik, 2009). Huge accumulation of
people in such events poses real life threats to
the people (Damian Roqueiro, 2006). A slight
mismanagement or panic may put at risks many
innocent lives. As a matter of fact, it is impera-
tive that such locations where people are likely
to be accumulated must be well designed or at

DOI: 10.4018/ijcvip.2013040101

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
2 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013
least well monitored for their safety. One way
of avoiding such a mishap is to count the num-
ber of people present there. Assuming that the
number of people should not exceed a particular
value, the system may raise an alarm if there
is an over accumulation of people. However,
counting people in crowded environments still
remain an elusive task.
People count can be used to monitor the
limit of occupancy ensuring that the people
count do not exceed a safety level for if the
number of people exceeds the capacity, people
may get crushed or the structure they are stand-
ing on may even collapse. Manual counting of
people has failed for different reasons. Firstly,
people tend to lose concentration when count-
ing and thus the counts are prone to errors.
Secondly, it is a very time consuming task as
well as difficult when peoples positions are
not fixed. Thus, there is a pressing need for an
automated system that can detect and count
people in crowded environments with reason-
able accuracy.
This paper proceeds as follows. Section 2
provides an in-depth literature review where
state-of-the-art approaches to estimate crowds
size are described. In Section 3, our proposed
systems are explained in detail. The solutions
proposed are then tested and evaluated in Section
4. Section 5 concludes the paper with a view
on the weaknesses of the proposed systems and
ideas for future work.
Figure 1. Features of Pedestrians: (a) original image, (b) foreground mask image (c) edge detec-
tion map, (d) the edge map after the AND operation between (b) and (c)
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
2. LITERATURE REVIEW
In Dan Kong (2006), a learning-based method
for counting pedestrians in crowds has been
proposed. The system uses edge orientation and
blob size histograms. This is done by applying
background subtraction and edge detection to
each frame. The features are then extracted
(Figure 1).
The points in the image and the point on
a 3D plane are related by a plane perspective
transformation called homography. Having ob-
tained the features, a homography is calculated
between the ground plane and the image plane
coordinates for the region of interest (ROI). For
feature normalization, a density map measuring
the relative size of persons and a global scale
measuring camera orientation have been used
(Figure 2).
Two training methods are used: Linear
Fitting and Neural Networks. There two types
of training because feature size and pedestrian
counts do not have a simple linear relationship.
As such, Neural Networks is used for the case of
occlusions in the crowd and counting of people
using Linear Fitting is used based upon edge
and blob histograms.
In Danny B. Yang (2003), the system
makes use of a group of image sensors con-
nected through a network to estimate crowd
density (Figure 3).
The first step is to compute this projection
from the silhouettes measured by the sensors
through background subtraction. The projection
is a set of polygons. Figure 4 shows some top
views of the resultant projections obtained:
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 3
Figure 2. Region of interest in the image and density map (weights)
The second step is to compute bounds to the
number of objects in each polygon. As objects
move, these bounds change and can be improved
over time. Polygons are pruned based on size
and temporal consistency. Polygons smaller
than the minimum object size and that appear
from nowhere are considered as phantoms and
are removed.
Also, objects may be completely occluded
by other objects. Therefore, the system keeps
track of a lower and upper bound of the number
of objects in each polygon with constraints for
the Upper and Lower bounds respectively. A
constraint for the upper bound is the area of
the polygon divided by the minimum object
size. For the lower bound, a polygon is said
to contain at least one object if there exists
a ray from a camera that intersects only that
polygon. Finally, a tree is used to keep track
of incoming and outgoing objects. This system
cannot be used in areas where there is a high
density of people.
Detecting skin-coloured pixels (Ahmed
Elgammal, 2009), has proven to be quite
Figure 3. Typical views of people from different sensors
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
challenging for many reasons. The appearance
of skin in an image may vary depending on the
illumination conditions (illumination geometry
and colour) where the image was captured (Jie
Yang, 1997). Humans are very good at identify-
ing object colours in a wide range of illumina-
tions. This is called colour constancy. Colour
constancy is in fact a mystery of perception.
Consequently, an important challenge in skin
detection is to represent the colour in a way that
is invariant to or at least insensitive to changes
in illumination (Gonzalez, 2005). Also, the
choice of colour space significantly influences
the performance of any skin detector. Another
challenge comes from the fact that many objects
in the real world might have skin-tone colours.
A few examples are wood, skin-coloured cloth-
ing, hair and sand. These real world objects
may affect the skin detector and will produce
many false detection in the background if the
environment is not adequately controlled.
Skin colour is usually used to track faces
and hands. Since it is invariant to orientation and
size, it provides an extra dimension compared
4 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013
Figure 4. Different persons arrangements can be consistent with a given image hull
to gray scale methods (Vladimir Vezhnevets,
2003). For instance, two objects of similar
gray tones might be very different in a colour
space. Skin detection is also relatively fast to
process. Normally, skin parts are detected us-
ing skin colour cues which vary from colour
spaces to colour spaces. The colour feature
is pixel-based and requires no spatial context
making it independent of the orientation and
size in contrast to other cues.
The estimation of the number of individu-
als in Marana (1999) is done by comparing
area occupation level with projected ones. It
is based on the Minkowski fractal dimension.
The latter has been widely used to character-
ize data texture in a large number of physical
and biological sciences. Here, it has been used
to illustrate levels of persons overcrowding
in crowd images. The first operation is edge
detection. Then, a binary image is generated.
Next, n dilations are computed from the binary
image with structuring elements of different
sizes, ranging from 1 to n. From these dila-
tions, the fractal dimension is estimated. The
fractal dimension is then used as a feature to
classify the image as belonging to one of the
five following density classes: VL very low
density; L low density; M moderate density;
H high density; or VH very high density.
Figure 5 shows corresponding density
classes on two different images of crowd density.
Left image: very low density. Right image: Very
high crowd density.
Using the Minkowski approach, sausages
are generated by using dilation operations with
structuring elements of different sizes 1, 3, 7,
and 11. Looking at the sausages in Figure 5,
it is possible to observe that the filled area of
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
the image grows as the size of the structuring
element also grows, until it reaches the total
area of the image.
Gwang-Gook Lee (2007) used a method
of estimating crowd size based on statistical
comparison of low-level image features and the
crowds size in the image. The system has three
main steps: preprocessing, feature extraction
and crowd size estimation.
Figure 6 shows the block diagram of the
proposed method. Preprocessing operations
deals with foreground segmentation and edge
detection. Foreground segmentation is achieved
by applying background subtraction while the
Canny edge detector is used for edge detection.
The feature normalization method consid-
ers camera projection and an elliptical human
model. In the elliptical human model, people
are represented by an ellipsoid with average
human height. By weighting each foreground
pixel with this scaling factor, the sizes of dif-
ferent blobs are standardised to the same size
irrespective of their locations (Figures 7 and 8).
If the height of an object is not more than
0.7 times an average human height, the latter
is not taken into consideration for counting.
Sirmacek and Reinartz (2011) uses airbone
images for crowd analysis. Even though the
images resolutions are not enough to see each
individual in details, there exists a change in
colour components in the place where a person
is located. A colour feature detection method
based on a probabilistic framework is used in
this work. Firstly, the local features are extracted
from illumination invariant chroma bands of
the image.
Secondly, assuming the extracted features
as observation points, a probability density
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 5

Figure 5. Crowd density classes

Figure 6. Steps for crowd size estimation

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
6 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013

Figure 7. Elliptical human mode

function using the Gaussian Kernel functions A thresholding operation is then applied to
with constant bandwidths is created. Next, an obtain the crowded regions. Finally, feature
adaptive method is established for choosing an locations at detected crowds are examined to
appropriate bandwidth for the Gaussian kernel. get the approximate number of people present

Figure 8. Counting decisions on the boundaries

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 7
in these crowded regions. Figure 9 shows a
group of three images illustrating how the crowd
density is obtained.
Tracking people in high density crowds
with lots of other objects is very difficult. Ali
(2009) presented an approach to detect and track
multiple individuals in highly dense crowds with
extreme occlusions. He used head detection
and trained a head classifier offline using Viola
and Jones AdaBoost cascade classification. To
acquire a good database of images, a significant
number of positive and negative images were
used. The resolutions of these samples were also
considered. The training samples were created
and converted to an XML file. The XML file is
used by an application to detect the heads in a
video. The total number of heads is then counted.
Figure 11 above shows the detection and
counting of people using a haar classifier trained
with positive and negative samples shown in
Figure 10a and Figure 10b.
Oliveira (2009) uses the Hue channel to
get the skin colour range. The author converted
the RGB values for each pixels in the image
to HSV values and uses a specific range of
values to highlight the skin by converting all
non-skin pixels to black and skin pixels to
white. Morphological operations were applied
to remove the noisy pixels for better perception
(Figures 12-15).
In an RGB-encoded digital image, the
colour of a pixel is represented by three values
ranging from 0 to 255, each represent one co-
lour channel Red, Green, or Blue. However,
Figure 9. Automatic crowd analysis from airborne images
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
the values are affected not only by the real
colour of the object, but also affected by the
condition of illumination. If we assume the
object is illuminated by a white light, then the
RGB encodings can be normalized as follows
(Haishi, 2009):
Intensity, I = (R + G + B) / 3
Normalised red, r = R / (R + G + B)
Normalised green, g = G / (R + G + B)
Normalised blue, b = B / (R + G + B)
Human skin can be detected by comparing
the pixels RGB colour falls in a specific defined
range provided a skin model is available. A cer-
tain range is defined for the following channels:
Minimum range: r >= 0.2 and r <= 0.46
and g <= 2 * r - 0.4
Maximum range: r > 0.46 and r <= 0.8
and g <= -r + 1
Using the ranges above, the result obtained
by applying the above threshold on the nor-
malised image is shown in Figure 16 (Figures
16 and 17).
The HSV colour space has three compo-
nents: the hue (H), the saturation (S) and the
brightness (V). Basically, HSV-type colour
spaces are deformations of the RGB colour
cube and they can be mapped from the RGB
space via a nonlinear transformation (K. N.
Plataniotis, 2000). One of the advantages of
these colour spaces in skin detection is that
8 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013

Figure 10. (a) Samples of positive images, and (b) samples of negative images

Figure 11. Detection of heads using Haar classifier

Figure 12. Original image

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 9

Figure 13. RGB pixels converted to HSV

Figure 14. Skin pixels only

Figure 15. Erosion followed by dilation

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
10 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013

Figure 16. Original image (left) and Normalised version (right)

they allow users to spontaneously identify the
boundary of the skin colour class in terms of
the hue and saturation (Maria Petrou, 2010).
There has been a consequent usage of this
particular colour space for skin detection by
several people, Shin (2002) and Albiol, Torres,
and Delp (2001).
Instruments (2008) showed how analysis
and processing of images are done on images so
as to extract valuable information about target
objects. A blob is a large binary object, where
all the object pixels are in the foreground are
white and all the other pixels are black in the
surrounding background. Yang (1994) showed
a comparison between area and perimeter and
circularity of blobs. However, the calculation
of perimeter is more complicated as the blobs
can be of irregular shapes. Assumptions are
made to reconstruct the contour in order to get
a better perimeter. Circularity of the blobs are
also taken into consideration where the result
for a circular object will be 1 and for all other
shapes it will be between 0 and 1.
Thou-Ho and Chao-Ho Chen (2007)
presented a way to count vehicles using blob
analysis in traffic surveillance system. The

Figure 17. Skin detected

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 11
different steps involved in implementing the
software were to segment the moving object, to
analyze and to track the blobs found. Meaning-
ful information was extracted from the analysis
of the vehicles blobs. Tracking was also done
by comparing the vehicles blobs on different
frames and calculating the minimal distance
between them. The density of the traffic flow was
also calculated by using the blobs information
in a predefined area. Shan Lu (2005) analysed
behavioural deception of people. The heads and
hands of people were identified through skin
colour detection and analysed based on the dif-
ferent movements and orientation of the heads
and hand blobs. The method was also tested on
different skin tones to measure its efficiency.
Cucchiara (2003) described how background
subtraction is used to extract features of moving
objects from videos in many applications such
as traffic monitoring, human motion capture
and video surveillance. The system uses colour
information for background subtraction and
shadow detection and tracks moving object and
shadows based on the processing obtained on
previous frames.
Hui-ying (2010) explained how to acquire
the ideal contours of vehicles for further pro-
cessing. Using OpenCV, the image is prepro-
cessed using morphological operations and a
thresholding operation is performed to convert
the image to a binary image. The contours are
then extracted from the image and thus the ve-
hicle is also detected and tracked. Malik (2001)
designed an algorithm to segment grayscale
images into distinguishable regions based on
their intensity levels. The cues of contours and
textures are analysed and are used to segment
the image.
3. PROPOSAL AND DESIGN
In this paper, we propose two algorithms to
automatically estimate the number of people in
a crowd. The following assumptions are made
about people in order to identify them: every
person is distinguishable visually from his/
her neighbour, the frontal or the profile view
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
is visible, the head or face is most visible than
any other parts of the body. The main steps of
both algorithms are shown in their respective
flowcharts provided below.
Figure 18 shows the main steps in the
blob detection system. Only blobs within a
certain height, width and number of pixels are
counted. Blobs which are too small or too big
are discarded. The number of people can then
be estimated via the number of blobs present.
Figure 19 shows the main steps in the contour
detection system. Only contours of a specific
length are considered. Contours which are too
short or too long are discarded. The number of
people can then be estimated via the number
of contours present.
After having converted an RGB image to
HSV format, we needed to find a good range
for skin values. For facial skin images, we
made the histograms of each of the three colour
channels, that is, H, S and V for a set of 120
images and we obtained the following ranges:
0 <= H <= 20, 30 <= S <= 154 and 83<= V <=
255 (Figure 20).
The next task in calculating the number of
people present from the binary image obtained.
We made use of blobs and contours separately
to count the number of people present by count-
ing the number of faces. However, the problem
that arose is that there are not only faces that
are being detected but also other skin regions
like hands, shoulders and legs. Therefore, we
had to define particular criteria for counting the
number of faces. For each skin blob or contour
present, we considered it as a face depending on
its weight, height and width. As such, arms and
legs are then discarded in the counting process
(Figures 21-22).
4. TESTING AND EVALUATION
As the estimations are based on head detection,
we assessed the performance of the two solu-
tions implemented on three different measures
namely: Total detection, Head detection and
False head detection.
12 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013

Figure 18. Blob detection algorithm

Total detection = head detection + false Completeness refers to the detection

head detection percentage.

False head detection refers to the head
detections done that are not considered as
people. False head detection occurs when the
system cannot distinguish between heads and
other objects.
Completeness = (total no. of head detec-
tions / actual number of people) x 100%
Correctness refers to the accuracy of the
system in identifying a person.

False head detection = Completeness - Correctness = (total no. of head detections

Correctness / actual number of people) x 100%

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 13

Figure 19. Contour detection algorithm

Figure 20. (a) RGB image, (b) HSV format, and (c) Binary image of skin detected

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
14 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013

Figure 21. Counting the number of detected blobs based on certain criteria

Figure 22. Counting the number of detected contours based on certain criteria

Table 1. Results of detection on 10 different images using blob detection

Image Total detection Head detection Actual people Total Head

count detection% Detection%
I1 115 107 127 90 84
I2 288 266 315 91 84
I3 103 89 122 84 72
I4 183 173 206 89 84
I5 138 125 152 91 82
I6 299 284 322 93 88
I7 129 115 148 87 78
I8 158 143 182 86 79
I9 119 112 136 88 82
I10 147 138 163 90 85

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013 15

Table 2. Results of detection on 10 different images using contour detection

Image Total detection Head detection Actual people Total Head

Count detection% detection%
I1 491 115 127 386 90
I2 1112 262 315 353 83
I3 726 103 122 595 84
I4 182 168 206 88 82
I5 399 132 152 262 86
I6 1051 258 322 326 80
I7 205 130 148 139 88
I8 518 137 182 285 75
I9 587 108 136 432 79
I10 632 145 163 388 89

The completeness and the correctness
percentages are tabulated and an average will
be used to contrast between the two solutions.
Table 1 shows the results from 10 different
images using blobs.
The average completeness rate is 89%,
average correctness rate is 82% and the average
false head detection rate is 7%. Table 2 shows the
results from 10 different images using contours.
The average completeness rate is 325%,
average correctness rate is 84% and the aver-
age false head detection rate is 241%. Thus,
both methods produce a reasonable detection
rate. However, the completeness factor for the
contour method still needs to be improved as
it is still far beyond a reasonable value. The
false detection rate for the blob solution is very
satisfactory. Completeness and correctness are
opposing measures. We can increase correct-
ness by decreasing completeness. This requires
selecting good threshold values for the different
operations involved.
crowd levels, judging the power and force of a
political party through estimates of its number
of supporters during a political meeting. We
have worked more towards making the system
correct rather than making the system complete.
However if some applications demand for more
accuracy, we can simply adjust the threshold
values and the parameters used in the different
functions in the application. However, there
are some limitations to our system. We have
assumed that the camera we have used to cap-
ture the crowd images uses a fixed focal length.
Thus, if the focal length changes continuously,
we will have to change the width, height and
number of pixels parameters constantly to get
a reasonable detection rate in our system. An-
other problem is the choice of the threshold for
extracting skin pixels as skins pixels vary from
dark pixels to bright pixels. In the future, we
intend to work with multiple-resolution images
and with crowds of varying densities.

5. CONCLUSION
In this paper, we have proposed two systems for
the automatic counting of people. We exploited
the fact that heads of people are more visible
in crowds. The system can be used in many
applications such as when monitoring safety
REFERENCES
Ahmed Elgammal, C. M. (2009). Skin detection - a
short tutorial.
Albiol, A., Torres, L., & Delp, E. (2001). Optimum
color spaces for skin detection. [Thessaloniki, Greece.
]. Image Processing, 1, 122124.

Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
16 International Journal of Computer Vision and Image Processing, 3(2), 1-16, April-June 2013
Ali, I. (2009). Detection and tracking of multiple
humans in high density crowds.
Bradski, G. A. K. (2008). Learning opencv.
Sebastopol, CA: OReilly Media.
Cucchiara, R. G. (2003). Detecting moving objects,
ghosts and shadows in video streams. Pattern Analysis
and Machine Intelligence, IEEE, 1337-1342.
Davies, A. C. (1995). Crowd monitoring using image
processing. IEE Electronic and Communications
Engineering Journal, 37-47.
Gonzalez, R. E. (2005). Digital image processing
(2nd ed.). New Delhi, India: Prentice Hall.
Gwang-Gook Lee, S.-H. S.-Y. (2007). Crowd size
estimation for video surveillance. In Proceedings
of the International Conference on Computing,
Communications and Control Technologies (ICCCT
2007), 2(196-199).
Haishi. (2009, december 22). Image processing.
Retrieved october 10, 2011, from http://haishibai.
blogspot.com/2009/12/image-processing-c-tutorial-
6-face.html
Instruments, N. (2008, May 15). Image analysis and
processing. Retrieved October 10, 2010, from http://
zone.ni.com/devzone/cda/tut/p/id/3470
Jie Yang, W. L. (1997). Skin-color modeling and
adaptation. School of Computer Science, Carnegie
Mellon University.
Jitendra Malik, S. B. (2001). Contour and
texture analysis for image segmentation.
International Journal of Computer Vision, 727.
doi:10.1023/A:1011174803800.
Lefloch, D. (2007). Real-time people counting system
using video camera.
Lim Eng Aik, Z. Z. (2009). Curve analysis for real-
time crowd estimation system. European Journal of
Scientific Research, 441-453.
Liu Hui-ying, W. X.-b. (2010). A vehicle contours
detection method based on OpenCV. Science
Technology and Engineering.
Copyright 2013, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
Luren Yang, F. A. (1994). Methods to estimate area
and perimeter of blob like objects. In Proceed-
ings of the 4 IAPR Workshop on Machine Vision
Applications.
A. N. Marana, L. d. (1999). Estimating crowd density
with Minkowski fractal dimension.
Moeslund, T. (2012). Introdction to video and image
processing. Springer-Verlag. doi:10.1007/978-1-
4471-2503-7.
Oliveira, A. C. (2009). Skin detection using HSV color
space. Brazil: Computation Institute Universidade
Federal Fluminense.
Petrou,M.C.P.(2010). Imageprocessingthefundamen-
tals. Singapore: Wiley. doi:10.1002/9781119994398.
Plataniotis, K. N. A. (2000). Colour image processing
and applications. Berlin, Germany: Springer-Verlag.
doi:10.1007/978-3-662-04186-4.
Roqueiro, D. V. A. (2006). Counting people using
video cameras. The International Journal of Parallel,
Emergent and Distributed Systems, 117.
Shan Lu, G. T. (2005). Blob analysis of the head
and hands: A method for deception detection. In
Proceedings of the 38th Annual Hawaii International
Conference on System Sciences, (p. 20).
Shin, M. C. (2002). Does colorspace transformation
make any difference on skin detection? Applications
of Computer Vision, (WACV 2002), Washington, DC
(pp. 275-279).
Sirmacek, B., & Reinartz, P. (2011). Automatic crowd
analysis from airborne images. In Proceedings of the
2011 5th International Conference on Recent Ad-
vances in Space Technologies (RAST) (pp. 116-120).
Thou-Ho & Chao-Ho Chen. Y.-F. L.-Y. (2007). Intel-
ligent vehicle counting method based on blob analysis
in traffic surveillance. In Proceedings of the Second
International Conference on Innovative Computing,
Information and Control (p. 238). Washington DC:
IEEE Computer Society.
Vezhnevets, V. S. (2003). A survey on pixel-based
skin color detection techniques. Moscow, Russia.
Yang, H. H.-B. (2003). Counting people in crowds
with a real-time network of simple image sensors.