See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330781235

Automatic Video Surveillance System for Pedestrian Crossing Using Digital

Image Processing

Article  in  Indian Journal of Science and Technology · January 2019

DOI: 10.17485/ijst/2019/v12i2/139859

CITATIONS READS
0 9,695

2 authors, including:

Amudha J
Dr. Mahalingam College of Engineering and Technology
19 PUBLICATIONS   14 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Deblurring of natural images View project

Deblurring of images in the presence of impulse noise View project

All content following this page was uploaded by Amudha J on 16 February 2019.

The user has requested enhancement of the downloaded file.

 ISSN (Print): 0974-6846
Indian Journal of Science and Technology, Vol 12(2), DOI: 10.17485/ijst/2019/v12i2/139859, January 2019 ISSN (Online): 0974-5645

Automatic Video Surveillance System for Pedestrian

Crossing Using Digital Image Processing
C. Thirumarai Selvi1 and J. Amudha2
1
Department of Electronics & Communication Engineering, Sri Krishna College of Engineering and Technology,
Coimbatore - 641008, Tamil Nadu, India; [email protected]
2
Department of Electrical and Electronics Engineering, Dr. Mahalingam College of Engineering and
Technology, Pollachi - 642003, Tamil Nadu, India; [email protected]

Abstract
Objective: This work proposes a novel technique to assist the physically challenged people to cross the road near the
traffic signal or in the pedestrian crossing with the help of automatic video surveillance. Methods/Statistical Analysis: As
per the Indian roadway corporation it is decided to minimum 1.8m (width) × 2.2m (height) is allocated for walking zone.
1800mm width is reserved for wheel chair movement in pedestrian. Detection of moving objects in a stable place needs
the high security level. Findings: Image processing segmentation algorithm plays an important role to track the moving
object in the fixed pedestrian crossing which can assist the physically challenged people or the elderly person. Various
morphological filtering operations enhance the quality of segmenting the moving person in the video. This methodology
employs histogram of Gaussian detection and object detection is done. The proposed work improves it results by 0.71%
accuracy. Applications/Improvements: This work helps the physically challenged to cross the pedestrian in a safe manner
and helps for autonomous vehicles.

Keywords: Image Processing, Morphological Operations, Pedestrian Crossing, SVM Classification, Video Surveillance

1. Introduction ­ urpose is called as video surveillance. The ­significance of

p
this system is to help the physically challenged people and
Surveillance cameras are placed in shopping mall, ATM elders to cross the traffic signal very safety. Video surveil-
machine, public places, industries, bank, educational lance importance is applicable in multiple fields known as
institution, hospitals, traffic signal and others. Static back missile tracking, security purpose, medical laparoscopy,
ground, moving objects, appearance of static and dynamic moving robot design during the building crash, road and
components of the scene and moving objects are the three forest accident prevention etc.
main components of the scene. With the advent of video Object tracking is the main sequence used for the
coding technologies, video surveillance is a blooming field video surveillance system. Figure 1 explains the general
for object tracking. Video surveillance with human is a block diagram of the video surveillance method. Frame
very time consuming process. Hence machine has to ana- difference method, optical flow method and background
lyze the video and extract the necessary information for subtraction methods are the significant methods applied
further application. Many soft computing algorithms have for object tracking. Optical flow method is little bit com-
the capacity to automatically detect various objects such plex and cannot be used in real time application. Among
as person, vehicle, animals, and equipment etc. Analyze the various methods, background subtraction method is
or observe a particular place for business and safety popularly efficient method selected for the work.

*Author for correspondence

 Automatic Video Surveillance System for Pedestrian Crossing Using Digital Image Processing

­reprocessing methods includes projection filtering,

p
monocular vision and contour information analysis.
This system gains high recall rate, robustness and preci-
sion. The given model can be working in the different
situation and light conditions. SVM machine learn-
ing algorithm is uses to avoid false identification. The
main benefit of this system is to detect perfectly even
in defilement and impairment traffic situation of roads.
Identification of pedestrians has been done in the work4
Figure 1.  Video surveillance general frame work. by combining motion information with image inten-
sity information. The detector employed with Adaboost
This study contributes algorithms for monitoring
neural network algorithm to detect a walking person as
pedestrian crossings and identifies the human or vehicle
well as motion appearing information. This study work
and other objects. With the trained template matching,
efficiently, even in the low resolution images such as rain
the activities are analyzed using image processing tech-
or snow environments. For preprocessing it utilizes a
niques for segmentation and behavior detection. To
simple rectangular filter with minimum extraction time
monitor this multiple video cameras are installed and data
of 0.25 ms to handle 360 × 240 image size.
are collected. And this system will work for automated
A novel method5 for detecting the pedestrian by
vehicle, and to prevent accidents for other vehicles. The
­combing HOG and optical flow method. Caltech and
study is organized as follows. Literature survey has been
ETHZ pedestrian benchmark images are applied for
presented in section 2. Section 3 presents the details of
analysis and provide low false positive rates. This auto-
existing method. The proposed methodology is discussed
matic pedestrian detection system also gives low average
in section 4. Section 5 gives the experimentation results
miss rates ranges. Two classifier models named as SVM
with comparison of existing method. Finally ­section 6
and Ada-boost methods are used for classification. A
summarizes the work performed.
data driven approach6 has been developed to detect
zebra crossings to examine the aerial images. HOG and
2.  Materials and Methods LBPH methods worked together with SVM algorithm is
employed here. This work can contribute a navigational
The research describes a zebra crossing detection1 system guidance for visually impaired people.
for operating an intelligent vehicle. This work adapts a
morphological filter followed by horizontal projection to
extract zebra crossing regions. Then self-similarity recog- 3.  Existing Work
nition method is employed to segment the zebra crossing
A part of the existing work was devoted to col-
region. This work achieves an accuracy of 98.5%. But this
lect relevant videos of pedestrians crossing near the
system, fails to monitor the traffic signal and its control
zebra crossings. This work7 investigates the features
operations. The average time taken to capture one frame
of pedestrian detection and behavior analysis. The
is about 57.24 Ms. The road scene segmentation process2
behavior analysis includes, eye gaze, face expression,
has developed using convolutional neural network. This
pose, body motion and hand movements. Advance
CNN extracts the road scene from the 3D scene. The
machine learning techniques are used to solve these
developed algorithm generates trained labels from gen-
tasks. Figure 2 explains the work flow of existing
eral image data set. From the trained labels, the real time
method.
road scenes can be compared and analyzed to understand
the road images. This work combines online images with
off line images. Compared to the baseline image analysis
methods, there is a 7% improvement of accuracy in the
noisy image.
The recent work3 for vehicle based mobile ­mapping
system for automatic pedestrian crossing. The Figure 2.  Block diagram of the existing work.

2 Vol 12 (2) | January 2019 | www.indjst.org Indian Journal of Science and Technology
 C. Thirumarai Selvi and J. Amudha

4. Methodology Table 2. List the objects used for classification

Pedestrian Classification Visually impaired
The proposed work in Figure 3 collects the data from recognition Blind
video and separates into different frames. For every frame Bicycle
the back ground can be subtracted from fore ground Vehicle
extraction and compares the foreground for every frame. Wheel chair
Histogram of gradient method is useful to create the his-
togram of different objects for different gradients. The Table 3. List of activities for second search
extracted foreground objects are classified using both
Pedestrian Detection recognition Action
Support vector machine and Kande algorithm using com-
Activity
puter vision. The classified objects based on activities it is Gesture
tracked for different velocity or speed. Gait
One of the efficient methods useful for object Behavior
­identification is Histogram of oriented gradients (HoG).
Using this local object identification and shape has been Table 4. Types of pedestrians
estimated using histogram. Initially the centered and Adults Disabled
horizontal gradients are recognized without smooth-
Elderly Injured
ing. Then the magnitude and gradients are computed.
Children Visually impaired
Table 1 lists the different features for identifying the
objects for segmentation. Table 2 is used for first level of Groups of children Bicyclists
object recognition for the first level of searching. After the People carrying heavy object
image classification, the second search for classification
the Table 3 lists the group of keywords. Table 4 gives the about 1.2 meter per second and for the fat walking it is
nature of pedestrians who may use the pedestrian. The given by 2 meter per second collected from Handbook
theoretical walking speed for the ordinary signaling is 048 from the NRPA for pedestrians. Machine learning
duplicates the physical activities of human behavior by
electronically perceiving and understanding an image.

4.1  Feature Extraction

Features are extracted as sub windows of a sample image.
Figure 3.  Block diagram of the proposed work. The base size of sub image is kept as 24 × 24. Four fea-
tures are scaled and shifted in all possible combinations.
Table 1. Case structure considered for the proposed Figure 3 represent the four different feature blocks. Figure
algorithm 4 specifies the subsamples considered for features.
Case structure
Feature Positive values
4.2  Algorithmic Steps of Feature Learning

Traffic flow Low, Normal, High • Learn a single simple classifier

Vehicle detected True, False • Classify the data
Speed of detected vehicle null, Slow, Normal, Fast
• Look at where it makes errors
• Reweight the data so that the inputs where we made
Number of positive Few, Normal, High
intentions errors get higher weight in the learning process
• Now learn a 2nd simple classifier on the weighted data
Time waited Short, Medium, Long
• Combine the 1st and 2nd classifier Figure 5 & 6 and
Type of pedestrian SlowGroup, Slow, Normal, Fast
weight the data according to where they make errors
Proposed transition time null, Extra-Extra time, Extra
• Learn a 3rd classifier Figure 7 on the weighted data and
(solution) time, Normal, Less time, Less-
less time so on until we learn T simple classifiers

Vol 12 (2) | January 2019 | www.indjst.org Indian Journal of Science and Technology 3
 Automatic Video Surveillance System for Pedestrian Crossing Using Digital Image Processing

• Final classifier Figure 8 is the combination of all T 5.  Experiment Results and
classifiers
• This procedure is called “Boosting” - works very well
Discussions
in practice. The experiment is simulated using Matlab and the
resulted are shown in Figure 9. Table 5 shows the com-
parative results of the existing method with the proposed
method. The results show that the proposed method
has improved accuracy by 0.71%. Also it shows there
is a reduce time for classification against the existing

Figure 6.  First classifier.

Figure 4.  Four different samples for feature extraction.

Figure 7.  2nd classifiers.

Figure 5.  Boosting. Figure 8.  Final Classifier learned by Boosting.

Table 5. Comparison between proposed and existing methods

Distance. Velocity Timing

Features Existing Proposed Existing Proposed Existing Proposed

Case1 96.71 97.03 211.54 213.31 0.456 0.441
Case2 97.23 97.54 214.63 216.22 0.456 0.441
Case3 97.66 97.86 211.62 211.65 0.456 0.441
Case4 99.02 99.50 209.10 210.49 0.456 0.441
Case5 99.11 99.83 205.92 205.98 0.456 0.441

4 Vol 12 (2) | January 2019 | www.indjst.org Indian Journal of Science and Technology
 C. Thirumarai Selvi and J. Amudha

Figure 9.  Object tracking results.

Figure 12.  Time between previous method and existing

Figure 10.  Distance between previous method and method.
existing method.
method. Figure 10-12 represents the overall comparison
of distance, velocity and computation speed.

6. Conclusion
In this work pedestrian walk detection system is
accomplished through multiple classifiers and the
­
detection rates are viable with some of the existing
methods to date in terms of both performance and
running time. The tracking method is responsible
for tracking the human activities by distinctive fea-
tures. Histogram of gradient method is used for object
Figure 11.  Velocity between previous method and existing
tracking together with its increase in speed. Hence
method.

Vol 12 (2) | January 2019 | www.indjst.org Indian Journal of Science and Technology 5
 Automatic Video Surveillance System for Pedestrian Crossing Using Digital Image Processing

this method can be used for various applications that 4. Viola P, Jones MJ, Snow D. Detecting pedestrians using
involve object tracking. patterns of motion and appearance. International Journal
of Computer Vision. 2005; 63(2):153-61. https://doi.
org/10.1007/s11263-005-6644-8
7. References 5. Ramzan H, Fatima B, Shahid AR, Ziauddin S, Safi AA.
Intelligent pedestrian detection using optical flow and
1. Wang C, Zhao C, Wang H. Self Similarity Based Zebra
HOG. (IJACSA) International Journal of Advanced
Crossing Detection for Intelligent Vehicle. The Open
Computer Science and Applications. 2016; 7(9):408-17.
Automation and Control Systems Journal. 2015; 7:974-86.
https://doi.org/10.14569/IJACSA.2016.070955
https://doi.org/10.2174/1874444301507010974
6. Koester D, Lunt B, Stiefelhagen R. Zebra crossing detec-
2. Alvarez JM, Gevers T, Le Cun Y, Lopez AM. Road scene
tion from aerial imagery across countries. Springer
segmentation form a single image. Springer-Verlag Berlin
International Publishing Switzerland. 2016; p. 1-8. https://
Heidelberg. 2012; p. 376-89.
doi.org/10.1007/978-3-319-41267-2_5
3. Liu X, Zhang Y, Li Q. Automatic pedestrian crossing
7. Detection and intention prediction of pedestrians in
­detection and impairment analysis based on mobile map-
zebra crossings. Available from: http://hh.diva-por-
ping system. ISPRS Annals of the Photogrammetry, Remote
tal.org/smash/record.jsf ?pid=diva2%3A1246982&ds
Sensing and Spatial Information Sciences. 2017; p. 18-22.
wid=1546940660347. Date accessed

6 Vol 12 (2) | January 2019 | www.indjst.org Indian Journal of Science and Technology

View publication stats