IRJET- Estimation of Crowd Count in a Heavily Occulated Regions

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 06 | June 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 2092
Estimation of Crowd Count in a Heavily Occulated Regions
Swathi D G 1, Jalaja G 2
1Student, Department of Computer Science and Engineering, BNMIT, Karnataka, India
2Associate Professor, Department of Computer Science and Engineering, BNMIT, Karnataka, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Crowd estimation is a challenging task of
accurately estimating number of people in a crowd region.
This paper aims to address crowd counting problem from the
perspective of two models i.e, body part map and structural
density map. The two models are created by combining the
information of pedestrian, their head and context structure.
Deep Convolutional neural networks and motion detection
method is used to count the number of people in the crowd
region, based on the pixel movement of the video frames. CNN
technique improves the efficiency of counting people in videos
and high accuracy is achieved.
Key Words: Crowd Counting, Deep Convolutional Neural
Networks, Motion Detection, Pedestrian Detection,
Crowd Estimation.
1.INTRODUCTION
Crowd estimation is thetask ofefficientlyestimatingnumber
of pedestrians in a dense region. Crowd counting has
harassed much curiosity from scientist due to the practical
stipulation like for controlling large number of pedestrians
and public security. Detection of a human is a basic issue in
video supervision systems. It is estimated that the world
population will be 11.2 billion in 2100years,whichisdouble
the current population of the world (7.4billon,2016).Dueto
rapidly growing population acrosstheworld,crowdanalysis
and crowd monitoring has become an important field for
research. Manually counting people in the dense crowded
areas user cannot estimate the accurate crowd count of the
pedestrians present in the area. To overcome this, a system
is developed to provide crowd count. Crowd count is any
dense scene is provided based on three key factors:
pedestrian, head and context structure, are planned as two
scene models. The first model is body-parts map, which is
obtained by finding the body parts of individual person in
dense scene and merging the segmentation mask. The
second model is structural-density map, which is created
based on shape of individual persons obtained from body-
parts map. Then result of two models are combined to
provide crowd count of the dense scene. There are several
applications of crowd counting some of them are listed
below: -
 Safety monitoring: - Video surveillance camera used in
public place for the safety and security of the people
may break down due to limitation in the algorithm
design of the system. In such scenarios, crowd counting
system can used for event detection, congestion control
and behavioral analysis.
 Intelligence gathering and analysis: - In malls and
airport, depending on the number of people entering or
length of queue the counters can be set up so that no
human resource is wasted.
 Designing a public place: - Crowd counting system can
be used to design public space like mall, stadium, rail
tracks etc.
2. RELATED WORK
Cross scene crowd estimation is a difficult task, where no
arduous data notations are required for estimating people
count of dense crowd scene. Deep convolutional neural
network (CNN) classifier is pre-trained to provide crowd
count of the dense scene-basedcrowddensity.Anewdataset
including 108 crowd images with 200000 head notations
was introduced to better evaluate accuracy of cross-scene
crowd estimation methods. To evaluate the efficiency and
reliability of the method experiment was held on already
existing datasets i.e, UCSD, UCF_CC_50 and WorldExpo’10
dataset. Cross-scene system fails to provide accurate count
of the dense crowd scene [1]. Pedestrian analysis is
challenging due to the gesture variation, obstruction,
appearance and background clutters.DeepDecompositional
network (DNN) classifier was used for parsing crowded
images into different human parts such as face, hairs, hands,
legs and body. Deep decompositional network together
estimates obstructed regions and body parts of person by
arranging three hiddenlayers:obstructionestimationlayers,
completion layers and decompositional layers. Pedestrian
parsing method by DNN provides betteraccuracythanstate-
of-art method on crowded images with or without
obstruction. The experiment was conducted on large
benchmark PPSS dataset for evaluating the efficiency and
reliability of pedestrian parsing method by DNN. The DNN
system fails to work efficiently in heavy crowded scene [2].
Global regression methods are used for mapping low level
features (texture, edge informationandsegmentationmask)
of humans to provide crowd count of the dense scene. The
system is evaluated over USCD dataset. The system ignores
the spatial information and body structure information of
pedestrian, thus fails to provide accurate crowd count of
crowded scene [3]. The head is the most visible part from
any crowded scene. The head detection is based on advance
method of boosted essential features. To reduce a search
region a novel point estimator base on gradient adjustment
features to identify region similar to the head region from

gray scale images. Head detector approach is evaluated on
PETS 2012 and Turin metro station datasets. Experiment on
these datasets gave good performance of head detector
approach for crowd counting. The Head detector method
fails to provide good performance in dense scene due to
obstruction short people where not visible [4]. Human
analysis and detection are the basic problem in any video
surveillance system. Shape based pedestrian detection uses
support vector machine (SVM) classifier for detecting
pedestrian in the crowded scenes. Support vector machine
classifier is pre-trained with few gestures of human to
separate human and no-human patterns. If test data gesture
matches in the train data then pedestrian will be included in
crowd count or else not. The Shape based method is
evaluated with three public datasets(INRIA,USC-BandMIT-
CBCL) and two benchmark datasets (Caviar and Munich
Airport). The Shape based system had many misdetections
due to the human pose estimation failure [5].
3. METHODOLOGY
Estimation of crowd count of any crowd video taken from
the crowded scene involves following image processing
steps as show in below figure 1.
Figure 1 System Architecture of Crowd counting system
3.1 Preprocessing
The crowd video is uploaded, then sequence of frames are
captured one by one from the video. The captured frame
contains distortion caused by the camera positing relativeto
object or position of objects in frames. Gaussian blur
technique is used to remove distortion and noise from the
frames, which results in blurring an input frame by gaussian
function.
3.2 Feature Extraction
Feature extraction is image processing techniques, where
the input raw data is reduced to manageable features for
processing. Motion detection is type of feature extraction
method where changing position of object is detected
relative to its surroundings. The preprocessed frames are
taken for detection humans in crowded video. The humans
are detected based on their movements relative to their
surroundings.
The motion detector algorithm used to detect the human
motion based on pixels movement includes 4 steps:
Step 1: Calculate the Difference between the
background_Frame and the current_Frame.
Step 2: The threshold value of the frame calculated in (Step
1) is used to filter the areas of motion.
Step 3: Resulting frame from (Step 2) is then highlighted in
the current_Frame to indicate areas of motion.
Step 4: Updated the background.
Figure 2 Motion Detector algorithm working
3.3 Deep Convolutional Neural Networks
Deep Convolutional neural network is type of neural
network used to classify the human and non-human objects
Vieo
Video
Preprocessing
Feature
Extraction
Deep CNN
Vieo
Crowd
Count

in the dense crowded scene. It is pre-trained with set of
videos to detect only human in different pose, then test data
is used to check the working of the neural network.
4. RESULT
Result of the proposed crowd counting system is shown in
figures below. For crowd counting system crowded video is
given as input and as the people starts moving in the video
crowd is obtained. The crowd count is estimated in both
direction i.e, number of people entering the in and out. The
threshold is set, when the person moves from thatthreshold
the crowd count is incremented by one else not. Proposed
system works well in both densely and low occulated
crowded regions.
Figure 3: Crowd counting method result for colored video
Figure 4: Crowd counting method result for grayscale
image
Figure 5: Crowd counting method result for human legs
(only)
5. CONCLUSION
In this paper, novel approach is presented to estimate the
crowd count in the crowd videos. The input video is pre-
processed to remove the distortion by using gaussian blur
technique. Gaussian blur method acts as low pass filter,
which remove the noise by blurring the inputframebyusing
gaussian function. The motion detector algorithm is used to
detect the motion of human with respect to background
structure, based on the human motion crowd count of given
video is estimated. CNNs classifier provides a better
performance than other neural networks classifiers. The
crowd count system provides a better performance even
with different human pose and illumination conditions. In
further, the crowdcountingsystemcanbeimplemented with
face recognition algorithm, so that each person is counted
only once.
REFERENCES
[1] Z. Lin and L. S. Davis, “Shape-based human detection and
segmentation via hierarchical part-template matching,”IEEE
Trans. Pattern Anal. Mach. Intell., vol. 32, no. 4, pp. 604–618,
Apr 2010.
[2] V. B. Subburaman, A. Descamps, and C. Carincotte,
“Counting people in the crowd using a generic head
detector,” in Proc. IEEE Conf. AVSS, pp. 450-470, Sep 2012.
[3] A. B. Chan and N. Vasconcelos, “Counting people with
low-level features and Bayesian regression,” IEEE Trans.
Image Process., vol. 21, no. 4, pp. 2160–2177, Apr 2012.
[4] P. Luo, X. Wang, and X. Tang, “Pedestrianparsingvia deep
decompositional network,” in Proc. IEEE ICCV, pp. 2648–
2655, Dec. 2013.
[5] C. Zhang, H. Li, X. Wang, and X. Yang, “Cross-scene crowd
counting via deep convolutional neural networks,” in Proc.
IEEE Conf. CVPR, pp. 833-841, Jun 2015.

[6] B. Wu and R. Nevatia, “Detection of multiple, partially
occluded humans in a single image by Bayesiancombination
of edgelet part detectors,” in Proc. IEEE ICCV, vol. 1. Oct.
2005, pp. 90–97.
[7] Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma, “Single-
imagecrowdcountingvia multi-columnconvolutional neural
network,” in Proc. IEEE Conf. CVPR, Jun. 2016, pp. 589–597.
[8] V. Lempitsky and A. Zisserman, “Learning to count
objects in images,” in Proc. Adv. NIPS, 2010, pp. 1324–1332.
[9] W. Ge and R. T. Collins, “Marked point processes for
crowd counting,” in Proc. IEEE Conf. CVPR, Jun. 2009, pp.
2913–2920.
[10] M. Wang and X. Wang, “Automatic adaptation of a
generic pedestrian detector to a specific traffic scene,” in
Proc. IEEE Conf. CVPR, Jun. 2011, pp. 3401–3408.
[11] A. N. Marana, L. F. Costa, R. A. Lotufo, and S. A. Velastin,
“On the efficacy of texture analysis for crowd monitoring,” in
Proc. IEEE SIBGRAPI, Oct. 1998, pp. 354–361.
[12] K. Chen, C. C. Loy, S. Gong, and T. Xiang, “Feature mining
for localised crowd counting,” inProc.BMVC,2012,vol.1.no.
2, p. 3.
[13] A. B. Chan and N. Vasconcelos, “Counting people with
low-level features and Bayesian regression,” IEEE Trans.
Image Process., vol. 21, no. 4, pp. 2160–2177, Apr. 2012.
[14] D. Kong, D. Gray, and H. Tao, “A viewpoint invariant
approach for crowd counting,” inProc.IEEEICPR,vol.3.Aug.
2006, pp. 1187–1190.
[15] A. B. Chan, Z.-S. J. Liang, and N. Vasconcelos, “Privacy
preserving crowd monitoring: Counting people without
people models or tracking,” in Proc. IEEE Conf. CVPR, Jun.
2008, pp. 1–7.
[16] D. Ryan, S. Denman, C. Fookes, and S. Sridharan, “Crowd
counting using multiple local features,” in Proc. IEEE DICTA,
Dec. 2009, pp. 81–88.
[17] N. Paragios and V. Ramesh, “A MRF-based approach for
real-time subway monitoring,” in Proc. IEEE Conf. CVPR, vol.
1. Jun. 2001, pp. I-1034–I-1040.
[18] M. von Borstel, M. Kandemir, P. Schmidt, M. K. Rao, K.
Rajamani, and F. A. Hamprecht, “Gaussian process density
counting from weak supervision,” in Proc. ECCV, 2016, pp.
365–380.
[19] M. Rodriguez, I. Laptev, J. Sivic, and J.-Y. Audibert,
“Density-aware person detection and trackingincrowds,” in
Proc. IEEE ICCV, Nov. 2011, pp. 2423–2430.
[20] Y. Wang, Y. X. Zou, J. Chen, X. Huang, and C. Cai,
“Example-based visual object counting with a sparsity
constraint,” in Proc. IEEE ICME, Jul. 2016, pp. 1–6.
[21] Y. Wang and Y. Zou, “Fast visual object counting via
example-based
density estimation,” in Proc. IEEE ICIP, Sep. 2016, pp. 3653–
3657.
[22] C. Arteta, V. Lempitsky, and A. Zisserman, “Counting in
the wild,” in Proc. ECCV, 2016, pp. 483–49.

IRJET- Estimation of Crowd Count in a Heavily Occulated Regions

More Related Content

What's hot (20)

Similar to IRJET- Estimation of Crowd Count in a Heavily Occulated Regions (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Estimation of Crowd Count in a Heavily Occulated Regions