International Journal of Computer Applications (0975 – 8887)

Volume 46– No.5, May 2012

An Insight into the Algorithms on Real-Time People

Tracking and Counting System

J. L. Raheja Pallab Jyoti Dutta

Machine Vision Lab, DSG Dept. of Electronics & Communication Engineering,
Council of Scientific & Industrial Research- Tezpur University
(CSIR-CEERI), Naapam, Tezpur,
Pilani, Rajasthan, India Assam, India

Sishir Kalita Solanki Lovendra

Dept. of Electronics & Communication Engineering, BKBIET, Pilani, India
Tezpur University
Naapam, Tezpur,
Assam, India

ABSTRACT
People tracking and counting system is being widely used in INPUT VIDEO
modern days in the video surveillance for security purpose. In SEQUENCES
the recent, many algorithms have been proposed and
developed in designing the system. But there always been
confusions of using these algorithms based on their limitations
MOVING OBJECT
and benefits. This paper provides a review on the methods
that are used in the people tracking and counting system DETECTION
considering their limitations, benefits, speed and accuracy. In
designing an efficient real time people tracking and counting
system, designers can be effectively guided by this review TRACKING
work.

Keywords
object detection, background modeling, tracking, counting, COUNTING
counting lines. ALGORITHMS
1. INTRODUCTION
People tracking and counting system is of great importance Fig. 1 People tracking and counting system
for surveillance purpose. At past, people involved in the
counting process and later with the development of electronic As seen from the above blocks, a short description of the
devices, the counting process was somewhat reliable. But both people counting system can be given as that the real-time
of them were not fully efficient, as there contained errors, and input video is taken from a camera. This video is further
also tracking could not be done with those processes. processes to detect to moving object. Then features are
extracted from the processed video for tracking and counting
So, to sort out those difficulties, computer vision techniques purpose.
have been used in the people counting system to give proper
counting process and tracking. People tracking and counting This review work is organized into five sections. Section 2
process involves many phrases such as detection of moving describes the background subtraction techniques. Various
objects, feature extraction, morphological processing, tracking tracking algorithms are described in Section 3. Section 4
and counting. For every process, there involves many describes the counting process. In Section 5, a proposed
algorithms to carry out the processes efficiently. Both the model has been described based on the advantages and
expert and the newcomer to this area can beconfused about limitations of these algorithms. Conclusive remarks are
the benefits and limitations of each step. This paper provides a described in Section 6.
thorough review of the main methods and their advantages 2. MOVING OBJECT DETECTION
and disadvantages based on speed and accuracy.
In a video sequence, one of the important steps is to
The design process of people counting system has been
detect the moving objectin the scene from static
described as the block given below. Different researchers use
cameras.Several methods for performing moving the object
different algorithms for their design. The blocks shown
detection have been proposed. All of these methods are used
consist of the algorithms that are widely used in the tracking
to estimate the background model effectively from the
and counting process.
temporal sequence of the frames.In this section, different
algorithms for detecting the moving objects have been
discussed. Some of the widely used object detection
techniques are as:

1
 International Journal of Computer Applications (0975 – 8887)
Volume 46– No.5, May 2012

Velastin et al. [4] as the background model. Even though the

2.1 Frame difference subsampling of n frames with respect to the original frame
2.2 Running Gaussian average rate by a factor of 10, Cucchiaraet al. [5] clarified that an
2.3 Temporal Median filter adequate background modeling can be provided by such
2.4 Mixture of Gaussians median values.In order to increase the stability of the
2.5 Kernel density estimation background model, [5] also proposed the computation of the
2.6 Eigen backgrounds median on a special set of values containing the last n
subsampled frames and w times the last computed values.
2.1 Frame Difference
Xiao Benxianet al. [1] used the frame difference method to 2.4 Mixture of Gaussians
detect the motion of the people moving in a video scene. His The mixture of Gaussians was proposed by Stauffer C et al.
works also stated that this method computes less cost and the [6].The detection or extraction of moving objects from the
simplest motion detection technique.It is the simplest method video scene is initial and most important step in the field of
for detection of moving object.It describes the detection of the video processing. Here, Gaussian mixture model is used to
foreground objects as the difference between the current describe a pixel of background. This statistical background
frame and an image of the static background of the scene. model allows multimodal background, thus providing robust
adaption against repetitive motion of scene elements, slow-
|𝑓𝑟𝑎𝑚𝑒𝑖 – 𝑏𝑎𝑐𝑘𝑔𝑟𝑜𝑢𝑛𝑑| > 𝑇𝑕 (1) moving objects, and introducing or removing objects from the
scene.
Where i represents the current ith frame, background is the
static background frame taken at the beginning and Th is the In the model the value of a particular pixel over time is seen as
threshold value. In this process, a first frame will be taken, a measurement, Xt of a stochastic variable.
which can be said as background. If the absolute difference is At any time, besides the current measurement of Xt, the
greater than threshold value, then the moving object in the history, Mt = {X1, X2,.……Xt-1}, of that particular pixel is
frame will be detected, otherwise it will be neglected. known [1]. Thus the recent history of a pixel can be modeled
using mixture of K Gaussian distributions. The probability to
observe the current background pixel Xt is the weighted sum
2.2 Running Gaussian Average of the K distributions:
Gaussians are very useful in determining the background
model. A proposal was given by Wren et al. in [2] to model 𝐾
𝑃 𝑋𝑡 = 𝑖=1 𝑤𝑖,𝑡 ∗ 𝜂 𝑋𝑡 , µ𝑖,𝑡 , ₡𝑖,𝑡 (5)
the background independently at each (i,j) pixel location. Here
a Gaussian probability density function (pdf) is being fitted
where K is the number of Gaussian distributions, wi,t is the
ideally on the last n pixel’s values. The following equation weight of the ith distribution at time t and ∑wi=1 , µi,t is the
shows the update equation of the background model at the mean value of the ith Gaussian at time t. ₡i,t is the covariance
appearance of the frame at time t, in where cumulative matrix of the ith Gaussian in the mixture at time t. η is the
average of the pixels is computed as:
probability density function of ith Gaussian, which is given by:
µ𝑡 = 𝛼𝐼𝑡 + (1 − 𝛼)µ𝑡−1 (2)
𝜂 𝑋𝑡 , µ𝑖,𝑡 , ₡𝑖,𝑡 = (1/((2∏)𝑛/2│₡𝑖 │1/2) ) ∗ 𝑒𝑥𝑝(−1/
whereIt is the pixel’s current value, µtthe previous average 2(𝑋𝑡 − µ𝑡 )₡𝑖 − 1(𝑋𝑡 − µ𝑡 ) (6)
andα is the learning parameter which can be considered as an
empirical weight often chosen as a tradeoff between stability K can be determined by computational and memory power.
and quick update. For quick learning, αshould be in-between Here we take three Gaussian distribution to describe a
0.5-1.0. As for each pixel, this consisted of two parameters (µt pixel.
,σt) instead of the buffer with the last n pixel values, where σ
is the standard deviation of the Gaussian pdf. The following
equation determines the condition for the foreground and To avoid the computational cost it is assumed that the
background. variation of the R, G and B color channel is same. So the
covariance matrix can be defined as:
|𝐼𝑡 − µ𝑡 | > 𝑘 𝜎𝑡 (3)
₡𝑖,𝑡 = 𝜎𝑖2 𝐼 (7)
If this inequality holds, then at each frame time t, the It pixel’s
value can be classified as foreground pixel, otherwise Itis The algorithm assigns one GMM to each pixel of the image
classified as background. Kollerel at.[3] modified the running and updates the model parameters (mean value and variance)
Gaussian average equation (3) as this equation unduely on-line. Every time when a new pixel is come it is checked
updated at the occurrence of foreground values. The modified with the already exist K Gaussian distributions. A match is
equation is: found if :
µ𝑡 = 𝑀µ𝑡−1 + (1 − 𝑀)( 𝛼 𝐼 + (1 − 𝛼)µ𝑡−1 ) (4)
│𝑋𝑖,𝑡 − µ𝑖,𝑡 │ < 2.5 ∗ 𝜎𝑖
where M represents a binary value. If M is 1, it corresponds to If none of the K distributions match the current pixel value,
the foreground value, and for background, it is 0. This the least probable distribution is replaced with a distribution
approach is also known as selectivebackground update. with the current value as its mean value, an initially high
2.3 Temporal Median Filter variance, and low prior weight. The weight is adjusted as
following.
The performance of the temporal average filter is shown as
better in the research works of many authors. A proposalto 𝑤𝑘,𝑡 = (1 − 𝛼)𝑤𝑘,𝑡−1 + 𝛼𝑀𝑘,𝑡 (8)
use the median value of the last n frames was put by Lo and

2
 International Journal of Computer Applications (0975 – 8887)
Volume 46– No.5, May 2012

Where α is the learning rate, Mk,t is 1 if match is found and 0 wheret is a given threshold. This approach can compensate for
if not match is found. The µ and σparameters for unmatched the changes such a shadows, climatic change, light intensity
distributions remain the same. For the match distributions etc. It can also offer less computational cost.
these parameters are updated as follows:

µ𝑡 = (1 − 𝜌) µ𝑡−1 + 𝛼𝑋𝑡 (9)𝜎𝑡2 = (1 −

3. TRACKING ALGORITHMS
2
Object tracking is a problem of estimating the positions and
𝜌) 𝜎𝑡−1 + 𝜌(𝑋𝑡 − µ𝑡 )𝑇 (𝑋𝑡 − µ𝑡 ) (10) other relevant information of moving objects in image
Where, 𝜌 = 𝛼𝜂(𝑋𝑡 │µ𝑡 , 𝜎𝑡 ) sequences. Tracking uses different features such as center of
gravity, color, HIS histogram etc. However, when occlusion
To decide which Gaussian of the GMM represents the occurs, it is very difficult to track multiple objects accurately.
background, they are sorted by the value of w/σ. This value People may create group when they are walking and may split
gets higher with less variance of the distribution and more from a heap.
times the distribution has been used. This leads to the
Gaussians of the GMM being sorted in a list where the first Tsong-Yi Chen et al. [10] used the center of gravity as a
more probably is background. Then the first B distributions feature to track down the moving objects in the succeeding
are chosen as the background model, where frames. According to their algorithm, they measured the
𝑏 Euler’s distance of the center of gravity in the adjacent frames
𝐵 = 𝑎𝑟𝑔𝑚𝑖𝑛𝑏 ( 𝑤𝑘 > 𝑇) as shown in the equation 13.
𝑘=1

T is a threshold on how much data should be accounted to bac

𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 = (𝑥𝑖 – 𝑥𝑖−1 )2 + (𝑦𝑖 – 𝑦𝑖−1 )2 (13)
kground. If T is chosen low usually a unimodal background w
ill be represented but with a higher T the probability of a multi
-modal background increases. For the same object to be tracked this distance between the
two successive frames must be minimum and less than a
threshold and a bounding box is drawn around the object.
2.5 Kernel Density Estimation
In statistics, kernel density estimation is a non-parametric way The Kalman filter can used to predict the future state of every
of estimating the probability density function of a random object in the next frame. The Kalman Filter is a statistical
variable. Kernel density estimation is a fundamental data method that involves an algorithm which provides an efficient
smoothing problem where inferences about the population are recursive approach in estimating the states of process by
made, based on a finite data sample. minimizing the mean of squared error. Kalman filter uses
elements of estimation theory to obtain the best unbiased
Elgammal et al. in [7] proposed a new way of modeling of the estimator of a state of a dynamic system using the previous
background based on Kernel Density Estimation (KDE) on measurement [11].BerilSirmacek et al. [12] used the Kalman
the last n background values. The sum of Gaussian Kernels filter for thetracking purpose. Kalman filter helps in the
centered in n background values, xt, is used to derive prediction of person in a frame of an image sequence based on
background pdf as given: the information derived from the previous frames. It uses the
𝑛
parameters of the previous frame and based on the
1
𝑃(𝑥𝑡 ) =
𝑛 𝑖=1
𝜂 (𝑥𝑡 − 𝑥𝑡 , 𝑡) (11) observation value of the current frame, it predicts the next
frame.
Piccardi et al. [8] stated that this model seems to be dealing
with a sum of Gaussians which is shown in equation (11). D. Comaniciu et al. [13] proposed an algorithm towards target
However, the main advantage is that each Gaussian describes representation and localization, the central component in
just one sample data, with n in the order of 100, and Z; is the visual tracking of non-rigid objects.Another tracking
same for all kernels. If background values are not known, algorithm used is Mean-shift based object tracking.
unclassified sample data can be used in their place; the initial DorinComaniciu et al. [14] used the Mean-shift tracking
inaccuracy will be recovered along model updates. Based on algorithm for the non-rigid objects by a single camera. To
(11), classification of xtas foreground can be straightforwardly find the target object similar to the present model, Mean-shift
stated if P(xt) <T. iterations are being employed. The similarity of these two
models is expressed by Bhattacharyya coefficient.Given the
predicted location of the target in the current frame and its
2.6 Eigen backgrounds uncertainty, the measurement task assumes the search of a
Nuria M. Oliver et al. in [9] built an eigen-space to detect the confidence region for the target candidate that is the most
moving object by modeling the background. In his work, the similar to the target model. The similarity measure that is
eigen-space was modeled by taking a sample of N images of p developed is based on color information.
pixels and computing the mean µb background images and
covariance matrix Cb. Using PCA, dimensionality reduction At first, target localization is done, then Weighted Histogram
of the covariance matrix is done to select the M largest Computation and then Distance Minimization using Mean-
eigenvalues matrix of ϕMb size M X p. Each input image Iican shift.The tracker employs two independent Kalman filters,
be projected to eigen spaces as which is expanded by the one for each direction x and y. The target motion is assumed
eigen backgrounds images Bi= ϕMbXi to model the static part to have a slightly changing velocity modelled by a zero mean,
of the scene. Hence, moving object was found out by low variance (0.01) white noise that affects the acceleration.
computing and thresholding the Euclidean distance between The tracking process consists in running for each frame the
the input image and projected image which can be shown as: mean shift based optimization which determines the
measurement vector and its uncertainty, followed by the
𝐷𝑖 = |𝐼𝑖 − 𝐵𝑖 | > 𝑡 (12) Kalman iteration which gives the predicted position of the

3
 International Journal of Computer Applications (0975 – 8887)
Volume 46– No.5, May 2012

target and a confidence region. These entities are used in turn

to initialize the mean shift optimization for the next frame. As described in [20] if the number of pixels of a particular
segmented image is greater than of a threshold value of pixels
Tao Liu et al. [15] proposed a mean shift object tracking of one person, then there is a possibility of containing more
algorithm. A novel weight to the given method is given, than one person. A bounding box is drawn for one person and
which improves the kernel function. The method is that the for two combined person, the box is divided into two
pixels which near the centre of object are given biggest rectangles.
weights, and the pixels which at the edge of the object are
given by exponent distribution as a result of occlusion. In
order to handle the occlusion, the occlusion detecting method
OUT LINE OUT LINE
based on sub-block detecting is also established. The novel
sub-block detecting algorithm is that the tracking window is
divided into two parts, including right and left, and the
similarity measure is calculated respectively. COUNTING ZONE

IN LINE IN LINE
4. COUNTING ALGORITHMS
To count the number of people entering or leaving a room,
Fig. 2 Fig. 3
different counting algorithms have been developed. For the
counting process to be accurate, different techniques are being
implemented. In this paper, some of the good counting
Equivalent Lines
algorithms are discussed.
Direction of Motion
In recent years, many computer vision based approaches for
people counting and tracking to deal various applications are
proposed. Researchers are trying to develop robust
background modeling algorithm to recognize the background
and track the moving object. Honglian Ma and Huchuan Lu et
Counting Zone
al. in [16] proposed a multiple people segmentation method
Fig. 4
based on the bi-directional projection histogram of grayscale
of frame differencing image. The insufficiency of this system
Fig.2: Single line counting technique Fig.3: Two line
is to only use the projection information of gray value as the
counting technique. Fig.4: Multiple lines method
detection criterion. The error in counting may occur, if the
projection information of other motion targets is similar to the
People whoever crossed this In Line as marked by the arrow
human targets. A.J.Schofield et al. [17] proposed a method to
head are counted as number of persons going in. Similarly,
count people in video images by using the neural networks.
this is for the people going out after crossing the Out Line in
They use RAM based neural network classifier to identify
the given direction. To detect and track moving people, Kim
section of background scene in each test image. Kenji Terada
et al. in [21] proposed a real-time scheme, where they used
et al. in [9] proposed a counting method in which they use
bounding box to enclose each person. The approximated
template process to detect the direction of the moving objects.
convex hull of each individual in the tracking area is obtained
From the moving direction information space time image
to provide more accurate tracking information.
generated. Then by counting the people data, the number of
incoming and outgoing person is counted. Chao-Ho (Thou-
Javier Barandiaran et al. [22] used more than two lines, i.e.
Ho) Chen et al. in [18] proposed a scheme to count the
multiple lines to count the number of people going in or
number of people entering and leaving a bus by using a
coming out within the range of the camera. The lines are
zenithal camera. They firstly divided the captured frame into
virtual and placed perpendicularly to the direction of the
many blocks then blocks with similar motion vectors regarded
motion of the people.The counting is performed by each line
as belonging to same object. Hartono Septian et al. in[19]
individually and the length must be bigger than half of a
proposed a method to count number of people, where they
person's width. Pixels are getting accumulated when someone
first detect the person as blobs and represents as binary masks
crosses the lines. To determine the direction of movement,
then a correlation based algorithm is applied to track the
people, Javier Barandiaran used Lucas-Kanade method
person in consecutive frames.
around the counting zone.
Researchers use techniques such as single line counting or 5. A PROPOSED MODEL
multiple lines counting for the counting process. Tsong-Yi All the algorithms for all the steps of a people counting
Chen et al. [20] proposed a counting algorithm where a single system have been discussed in above. All these algorithms are
camera is mounted at the top of the room. The algorithm used widely used. But all these algorithms are not used considering
a single line in the screen as the counting zone. The persons their advantages, speed and limitations. Based on these
crossing this line are counted as going in or coming out of the properties, a model is proposed for the designing of people
room. He took the object’s perimeter and area as the most tracking and counting system.
attainable features, for single image, for counting purpose.
They have also put a suggestion for distinguishing the In the moving object detection algorithm, it cannot be said
overlapping persons when it occurs in the frame. Analyzing that all the algorithms are good. The simplest of them is frame
the processed image, many features, such as the height, width, difference method. One of the main reasons is that of its
area, gravity center, contour, shape and pixels’ number, can modest computational load. Another is that the background
be obtained. From all these, only a few features of interest are model is highly adaptive to the changes in background, as it
selected to analysis. depends solely on the previous frame, which makes it faster

4
 International Journal of Computer Applications (0975 – 8887)
Volume 46– No.5, May 2012

than any other methods. The frame difference method has also [6] Stauffer C, Grimson W. E. L., “Adaptive Background
the capability to subtract out extraneous background noise Mixture Models for Real-Time Tracking”, Proceedings
(such as waving of leaves), much better than the more of conference on Computer Vision and Pattern
complex approximate median and mixture of Gaussians Recognition (Cat. No PR00149). IEEE Computer Society
methods. Vol. 2, pp. 246-252, June 1999.

For the tracking systems that are discussed, Kalman filter [7] A. Elgammal, D. Harwood and L. S. Davis, “Non-
tracking is one of the simplest and fastest tracking algorithms Parametric Model for Background Subtraction”, Proc.
that is used in this system. The advantage of Kalman filter- ECCV 2000, pp. 751-767, June 2000.
based techniques lies in their efficiency and in the high [8] M. Piccardi, “Background subtraction techniques: a
accuracy that can be obtained. It is predictive and adaptive as review”, IEEE International conference on Systems, man
it looks forward with an estimate of the covariance and mean and Cybernetics, pp. 3099-3104, 2004.
of the time series one step into the future, whereas in mean-
shift algorithm, the implementation in software takes a long [9] N. M. Oliver, B. Rosario and A. P. Pentland, “A
time which is about 15 fps. Bayesian Computer Vision System for Modeling Human
Interactions”, IEEE Trans. on Pattern Anal. and Machine
For the counting process, the two line counting process can be Intell., Col. 25, No. 11, pp. 1499-1504, 2003.
found more reliable compared to others. In this system, two [10] Tsong-Yi Chen, Chao-Ho Chen, Da-Jinn Wang and
virtual lines are drawn such that the distance from bottom Tsang-Jie Chen, “Real-Time Counting Method for a
edge to the IN LINE and from top edge to the OUT LINE is Crowd of Moving People”, Sixth International
greater than the bounding box of the person entering or Conferenceon Intelligent Information Hiding and
leaving the room. Using all these algorithms, a real time Multimedia Signal Processing, pp. 643-646, 2010.
people counting system can be designed based on their
accuracy, speed etc. [11] Raul Rojas, “The Kalman Filter”, available on
www.robocup.mi.fu-berlin.de/buch/kalman.pdf, pp. 1-7.
6. CONCLUSION [12] BerilSirmacek and Peter Reinartz, “KALMAN FILTER
In this paper, a review of the algorithms that are widely used BASED FEATURE ANALYSIS FOR TRACKING
in real-time people tracking and counting system have been PEOPLE FROMAIRBORNE IMAGES”, Proc. ISPRS
described. This can help many of researchers to select the XXXVIII.
appropriate algorithm in designing the system in real-time
field based on the advantages and limitations. We have also [13] D. Comaniciu, V.Ramesh and P.Meer, “Kernel Based
proposed a system for real time tracking and counting which Object Tracking”,IEEE Trans. Pattern Analysis and
may be useful to the researchers. Based on the proposed Machine Intelligence, Vol. 25, No. 5, 564-575, 2003.
system given above, it gives less computational time and also [14] D.Comaniciu and P.Meer, “Mean-shift Analysis and
tracking can be done effectively in real-time purpose. Applications”, The Proc. of the 7th IEEE International
Conference on Computer Vision, Vo. 2, pp. 1197-1203,
7. ACKNOWLEDGEMENT 1999.
This research is being carried out at Central Electronics [15] Tao Liu, Xiaping Cheng, “Improved Mean Shift
Engineering Research Institute (CEERI), Pilani, India as part Algorithm for Moving Object Tracking”, 2nd
of our project “Supra Institutional Project on Technology International Conference on Computer Engineering and
Development for smart systems‟. Authors would like to thank Technology(ICCET), Vol. 1, pp. 575-578, 2010.
Director, CEERI for his active encouragement and support
[16] Honglian Ma and Huchuan Lu Mingxiu Zhang, “A Real-
8. REFERENCES time Effective System for Tracking Passing People
[1] Xiao Benxian, Lu Cheng, Chen Hao, Yu Yanfeng and Using a Single Camera”, Proc. Of the 7th World Congress
Chen Rongbao, “Moving object detection and on Intelligent Control and Automation, June 25-27, 2008,
recognition based on the frame difference algorithm and Chongging, China, pp. 6173-6177.
moment invariant features”, 27th Chinese Control [17] A. J. Schofield, P. A. Mehta, T. J. Stonham, “A System
Conference (CCC), pp. 578-581, 2008. for Counting People in Video images using Neural
[2] C. Wren, A. Azarbayejani, T. Darrell, and A. P. Networks to identify the Background scene”, Journal of
Pentland, “Pfinder: Real-time Tracking of the Human Pattern Recognition, Vol. 29, Issue no. 8, pp. 1421-1428,
Body”, IEEE Trans. on Pattern Anal. And Machine 1996.
Intell., Vol. 19, No. 7, pp. 780-785, 1997. [18] C.H. Chen,Y.C. Chang, T.Y. Chen, D.J. Wang,“People
[3] D. Koller, J. Weber, T. Huang, J. Malik, G. Ogasawara, Counting System for Getting In/Out of a Bus Based on
B.Rao and S. Russell, “Towards Robust AutomaicTraffic Video Processing” IEEE Computer Society, Eighth
Scene Analysis in Real-Time”, Proc. ICPR’94, pp. 126- International Conference on Intelligent Systems Design
131, Nov. 1994. and Applications, pp. 565-569, 2008.

[4] B.P.L. Lo and S.A. Velastin, “Automatic Congestion [19] S. Hartono, T. Ji, T. Yap-Peng, “People Counting by
Detection System for Underground Platforms”, Proc. Video Segmentation and Tracking”, 9th international
ISIMP 2001, pp. 158-161, May 2001. Conference on Control, Automation, Robotics and
Vision, pp. 1-4, 2006.
[5] R. Cucchiara, C. Grana, M. Piccardi and A. Prati,
“Detecting Moving Objects, Ghosts and Shadows in [20] Tsong-Yi Chen, Chao-Ho Chen, Da-Jinn Wang and
Video Streams”, IEEE Trans. On Pattern Anal. and Tsang-Jie Chen, “Real-Time Counting Method for a
Machine Intell., Vol. 25, No. 10, pp. 1337-1442, 2003. Crowd of Moving People”, Sixth International

5
 International Journal of Computer Applications (0975 – 8887)
Volume 46– No.5, May 2012

Conference on Intelligent Information Hiding and [22] Javier Barandiaran, Berta Murguia and Fernando Boto,
Multimedia Signal Processing, pp. 643-646, 2010. “Real-Time People Counting Using Multiple Lines”,
IEEE Ninth International Workshop on Image Analysis
[21] J. W. Kim, K. S. Choi ,B. D. Choi, S. J.Ko, “Real-time for Multimedia Interactive Services, pp. 159-162, 2008.
Vision-based people counting system for security door”,
International Technical Conference on Circuits/Systems
Computers and Communications, pp. 1416-1419, 2002.