2021 International Symposium of Asian Control Association on

Intelligent Robotics and Industrial Automation (IRIA)

Indian Institute of Technology, Goa, India. 20-22 September 2021
2021 International Symposium of Asian Control Association on Intelligent Robotics and Industrial Automation (IRIA) | 978-1-6654-3323-5/21/$31.00 ©2021 IEEE | DOI: 10.1109/IRIA53009.2021.9588734

Detection of Normal and Abnormal Conditions for

Boundary Surveillance using Unmanned Aerial Vehicle
Kamal Sandeep Karreddula Alok Kanti Deb
Department of Electrical Engineering Department of Electrical Engineering
Indian Institute of Technology, Kharagpur, India Indian Institute of Technology, Kharagpur, India
[email protected] [email protected]

Abstract-There are several advanced methods to detect movie data with the help of Okutama-action dataset
objects, humans, etc., which requires large dataset for which comprises of 12 categories of human actions.
training. In this paper, a difference of difference (DoD) Ichraf Lahouli et al. [10], proposed a method for the
technique has been proposed to detect normal and detection of humans using thermal images with the
abnormal conditions with minimal dataset during help of thermal database.
surveillance. This dataset has been used for setting a One thing can be observed in advanced techniques,
threshold value to detect condition without training. large database has been required for detection of
Captured images have been compared with created objects, humans, etc. Background subtraction
dataset toget net difference value and this value has been
technique is widely used to detect moving objects [11].
compared with the threshold value to detect normal and
abnormal conditions. This technique can be used at no- Most of the background subtraction based techniques
man zone areas, boundaries and security borders, and have been performed well when camera is in static
also it can be used as a front-liner along with any position for monitoring [12]-[14].
advanced technique for surveillance application. This Z. Zaheer et al. [15] described an aerial
method has been simulated on real images using surveillance system using manually operated UAVs
MATLAB and conducted practical experiment using from the ground control station (GCS) with the help of
quad rotor, RPi and ROS. live stream. D. Avola et al. [16] proposed a technique
for aerial surveillance to identify changes in
Keywords: Object detection, surveillance, autonomous UAV
environment based on histogram equalization and
RGB-Local Binary Pattern (RGB-LBP) operator. N.
I. INTRODUCTION H. Motlagh et al. [17] presented a high level view of a
UAV based IoT platform and demonstrated
Unmanned Aerial Vehicles (UAVs) have become surveillance on crowd using face recognition
increasingly popular in both military and civilian technique. These results have been demonstrated the
applications. Time to time monitoring of the border is energy saving benefit of offloading computation. H.
very important to stop intruders or unauthorised Teng et al. [18] developed optimal trajectories based
activities near boundaries. So, autonomous border on particle swarm optimization for multiple
surveillance using UAVs will be very effective and it monitoring drones to monitor a certain operational
will also reduce the presence of unauthorised objects area, and detects illegal drones.
or activities near boundaries. Many different approaches and techniques have
In recent years, advanced techniques such as CNN been described by researchers, few of these have been
[1]-[3], RetinaNet [4], SSD [5], YOLO [6], HOG with performed in off-line mode, and few of these have
SVM [7] etc., have been used for surveillance been performed in online mode by transmitting the
applications to detect people. These have been camera data to GCS for processing, and few of these
required training on large dataset to achieve good have been performed in both off-line and online. But
performance. These techniques may not perform well all of these have been processed at the GCS by sending
in other domains, if the related domain image dataset data to high level processors to utilize advance
has not been included in the training. So, it requires techniques [19], [20].
additional training on the selected image dataset. Surveillance task consists of several stages like
Sasa S. and Marina I [8], discussed automatic processing, detection and communication etc. This
person detection using Deep CNN for search and work has been focused on detection of conditions. It is
rescue operation application. To test this application, online process without any human operator. So, it can
Deep CNN has been trained with related domain be considered as autonomous online surveillance
image set (i.e., bird view image set). Fan Yang et al. technique.
[9] successfully used atomic visual action (AVA) This proposed DoD technique can also be used in
detection to recognize “who is doing what?” in the fixed position monitoring system as well as in UAV
based surveillance system. This paper has been

978-1-6654-3323-5/21/$31.00 ©2021 IEEE 50

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 discussed simulation and experimental results of UAV
based surveillance using DoD technique. Advanced
techniques can be used for surveillance application but
these are heavy and consume more power during
continuous surveillance. So, proposed method can be
used as detector to detect normal and abnormal
Fig. 2 Waypoints and surveillance/reference points marking
conditions during surveillance at no-man zone areas
and boundaries. This method can also be used as a
Length of 𝑖𝑡ℎ surveillance point is 𝐿𝑆𝑖 and edges are
front-liner along with advanced technique in
surveillance which helps in reducing the power 𝑆𝑖𝑒1 and 𝑆𝑖𝑒2 . Here, 𝑑𝑆𝑖𝑒1 and 𝑑𝑆𝑖𝑒2 are distances of 𝑖𝑡ℎ
consumption. Different shapes of boundary have been surveillance point with respect to waypoint 𝑊1 .
converted to straight line boundaries by cutting into Surveillance point terminology has been illustrated in
pieces from corner to corner. Fig. 3.
In section II, methodology of the proposed method
has been discussed for surveillance application. In
section III, flow charts have been discussed to use the
DoD method as a detector and as a front-liner along
with HOG – SVM classifier. In section IV,
performance of the method has been discussed using
simulation and experimentation results. In section V, Fig. 3 Surveillance point terminology
concluding remarks have been discussed.
Proposed method has been used to identify
II. DETECTING NORMAL AND ABNORMAL abnormal condition without heavy dataset and training.
CONDITION Only two sets of images have been required at each
Presence of any object or a person at restricted surveillance point and these images have been
place may lead to unexpected and dangerous situation. captured at different flight times in normal condition to
So, these situations can be treated as abnormal create database (𝐼𝐷𝑆). Database – I (DB – I) and
condition of the boundary. UAV needs to move Database – II (DB – II) consists of a single image-set
different points during boundary surveillance. of each surveillance point under normal condition, and
Different shapes of surveillance boundary have been they have been captured during test flight – I and test
illustrated in Fig. 1. Different point terminology has flight – II respectively.
been used in this method to perform surveillance; they Let’s consider a straight line boundary consists of
are waypoints and surveillance/reference points. surveillance points (𝑁𝑆).
DB – I is as follow:
𝐼𝐷1 𝑆𝑖 = [𝐼𝐷1 𝑆1 , 𝐼𝐷1 𝑆2 , … , 𝐼𝐷1 𝑆𝑁𝑆 ] (1)
and DB – II is as follow:
𝐼𝐷2 𝑆𝑖 = [𝐼𝐷2 𝑆1 , 𝐼𝐷2 𝑆2 , … , 𝐼𝐷2 𝑆𝑁𝑆 ] (2)
In 𝐼𝐷1 𝑆𝑖 , 𝐼𝐷 indicates image database and
a b c d e
subscript gives the database number and 𝑆𝑖 indicates
Fig. 1 Types of boundaries, a) Straight line, b) Curved line, surveillance point or position number.
c) Rectangle, d) Pentagon, e) Hexagon Database 𝐼𝐷1 𝑆𝑖 and 𝐼𝐷2 𝑆𝑖 have been created during
test flight – I and flight – II respectively under normal
Waypoint is the point where UAV changes its path condition. Captured RGB images during surveillance
direction as per requirement. Waypoints have been have been converted into binary images (𝐼𝑏𝐷𝑆).
selected in such way that, UAVs should cover the Following (3) and (4) represents binary image
given boundary effectively. Surveillance/ reference database ,
point is the point where UAV holds its position and 𝐼𝑏𝐷1 𝑆𝑖 = [𝐼𝑏𝐷1 𝑆1 , 𝐼𝑏𝐷1 𝑆2 , … , 𝐼𝑏𝐷1 𝑆𝑁𝑆 ] (3)
captures image for detection process. UAVs should fly 𝐼𝑏𝐷2 𝑆𝑖 = [𝐼𝑏𝐷2 𝑆1 , 𝐼𝑏𝐷2 𝑆2 , … , 𝐼𝑏𝐷2 𝑆𝑁𝑆 ] (4)
through each surveillance point to perform the task. Difference of pixel magnitude values have been
Captured image should cover the allotted boundary obtained using the binary image database at each S
area between successive surveillance points. point by comparing images of DB – I and DB – II.
Waypoints and surveillance points have been Net Difference value at surveillance (𝑁𝐷𝑆)
illustrated in Fig. 2. point 𝑆𝑖 is given by,

51
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 𝑁𝐷𝑆𝑖 = ∑𝑙 ∑𝑚(|𝐼𝑏𝐷1 𝑆𝑖 (𝑙, 𝑚) − 𝐼𝑏𝐷2 𝑆𝑖 (𝑙, 𝑚)|) (5) 𝑇𝑁𝐷 𝑆𝑛 = 𝑁𝐷𝑆𝑛 + 𝜖𝑛
where, n = Ns, 𝑖 = 1,2. . 𝑛, and 𝑙 is horizontal and 𝑚 is where, 𝜖1 , 𝜖2 , … , 𝜖𝑛 are tolerance values to
vertical maximum pixel size of an image. corresponding net difference values.
Depends on the requirement, net difference values General equation will be represented as follows,
can be different or can be same value for S points. If 𝑇𝑁𝐷 𝑆𝑖 = 𝑁𝐷𝑆𝑖 + 𝜖𝑖 𝑓𝑜𝑟 𝑖 = 1,2, . . , 𝑛 (12)
all surveillance areas are similar type then, it can be For normal condition,
same value otherwise can be different value. 𝑁𝑆𝑆𝑛 ≤ 𝑇𝑁𝐷 (13)
For similar type of surveillance areas, For abnormal condition,
𝑁𝐷𝑆1 = 𝑁𝐷𝑆2 = 𝑁𝐷𝑆3 = ⋯ = 𝑁𝐷𝑆𝑛 = 𝑁𝐷𝑆 (6) 𝑁𝑆𝑆𝑛 > 𝑇𝑁𝐷 (14)
For different type of the surveillance areas, Threshold value is same for all S point in (13), (14).
𝑁𝐷𝑆1 ≠ 𝑁𝐷𝑆2 ≠ 𝑁𝐷𝑆3 ≠ ⋯ ≠ 𝑁𝐷𝑆𝑛 (7) For normal condition,
UAV based autonomous surveillance flights will be 𝑁𝑆𝑆𝑛 ≤ 𝑇𝑁𝐷 𝑆𝑛 (15)
started after creating the dataset and setting the net For abnormal condition,
difference value for S points. During surveillance 𝑁𝑆𝑆𝑛 > 𝑇𝑁𝐷 𝑆𝑛 (16)
flight, UAV holds its position at each S point and Threshold value is not same for all S point in (15), (16).
captures image (𝐼𝑆𝑆). After capturing image, it will So, the above equations (15), (16) can be treated as
start processing to detect the condition with the help of generalized equations for detecting normal and
on-board companion computer. UAV captures image abnormal conditions using the proposed method.
at 𝑆1 during 1st surveillance i.e., 𝐼𝑆1 𝑆1 which Here in this method, threshold value has been
represents the image of surveillance flight – I at S considered based on database image difference, and
point – 1. Captured RGB image during surveillance at net difference value has been calculated from
S point will be converted into binary (𝐼𝑏𝑆𝑆). Binary surveillance and database images. These have been
image of surveillance flight – I at S point – 1 (𝐼𝑏𝑆1 𝑆1 ) checked for deviation to detect normal and abnormal
will be compared with either binary image of DB – I condition. So, the proposed method can be called as
or with binary image of DB – II or both. After difference of difference method (DoD).
comparing with database image at S point – 1, net Important note is that Difference in difference (D
pixel magnitude during surveillance at S point (𝑁𝑆𝑆) method or DD method or D estimator or DD estimator
has been obtained. This net pixel magnitude value will [21]) is different than the proposed DoD method.
be compared with threshold value 𝑇𝑁𝐷 . This threshold
value has been finalised after getting the net difference III. FLOW CHART OF DoD METHOD
value from comparison of the database image.
Net pixel magnitude at S point – n during 1st Difference of difference (DoD) method has been
surveillance has been computed using the follow used to detect normal and abnormal conditions, and
equation, also used as a front-liner along with the advanced
𝑁𝑆1 𝑆𝑛 = ∑𝑙 ∑𝑚(|𝐼𝑏𝑆1 𝑆𝑛 (𝑙, 𝑚) − 𝐼𝑏𝐷1 𝑆𝑛 (𝑙, 𝑚)|) (8) technique to identify the desired objective. Below Fig.
Without counting number of surveillance and database 4 illustrates the DoD method flow chart to detect
number, the above (8) can be represented as normal and abnormal conditions alone and Fig. 5
𝑁𝑆𝑆𝑛 = ∑𝑙 ∑𝑚(|𝐼𝑏𝑆𝑆𝑛 (𝑙, 𝑚) − 𝐼𝑏𝐷𝑆𝑛 (𝑙, 𝑚)|) (9) illustrates the DoD method flowchart as a front-liner
Threshold value at 𝑆𝑛 is 𝑇𝑁𝐷 𝑆𝑛 . If all values are along with any advanced technique.
same, then Threshold value will be 𝑇𝑁𝐷 for each S Flowchart – I has been followed by the DoD
point. In ideal case, Threshold value will be equal to method to detect normal and abnormal conditions as a
the net difference value of database images. But in detector. After creating the required database and
practical cases, Threshold value will equal to the net setting threshold value for S points, autonomous
difference value of database images with some surveillance has been started using UAV. Images have
additional tolerance value 𝜖. been captured using UAV at each S point to start the
For ideal cases, process. If net difference value is greater than
𝑇𝑁𝐷 𝑆1 = 𝑁𝐷𝑆1 , 𝑇𝑁𝐷 𝑆2 = 𝑁𝐷𝑆2 , … , 𝑇𝑁𝐷 𝑆𝑛 = 𝑁𝐷𝑆𝑛 threshold value, then it will be considered as abnormal
If all threshold values are same, condition otherwise it will be considered as normal
𝑇𝑁𝐷 = 𝑁𝐷𝑆 (10) condition. For abnormal condition, UAV will hold its
For practical case, if all threshold values are same: position and hovers at S point only. It will capture
𝑇𝑁𝐷 = 𝑁𝐷𝑆 + 𝜖 (11) images to do detection process till it get clearance
where, 𝜖 is tolerance value, it can be 5 to 10% of net from the ground control station (GCS) or till it detect
difference value. normal condition. If it identifies normal condition,
For practical case, if threshold values are not same: UAV moves to next S point and repeats the same
𝑇𝑁𝐷 𝑆1 = 𝑁𝐷𝑆1 + 𝜖1 ; 𝑇𝑁𝐷 𝑆2 = 𝑁𝐷𝑆2 + 𝜖2 ; … ; process.

52
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 Flowchart – II has been followed by the DoD Flowchart – II:
method as a front-liner along with advanced technique
in identification process. In this, images have been
captured to detect normal and abnormal condition at S
point and moved to next till it detects abnormal
condition. For abnormal condition, advanced
technique has been used to detect exact classification
as per the given training. If it detects presence of a
person or object as per the training at restricted area,
then UAV will hold its position and captures image
and it repeats the same process till it detects normal
condition or till it get clearance from the ground
control station (GCS). This method helps in reducing
power consumption for advanced technique using the
DoD method as a front-liner during surveillance.

IV. RESULTS AND DISCUSSION

In this section, the DoD method has been simulated

on real images. Fig. 6 to Fig. 8 and Table – I shows
results and performance of the method on real images.
Practical experiments have been conducted using Fig. 5 DoD method flowchart as a front-liner along with advanced
quad-copter, Raspberry Pi (RPi) as companion techniques
computer and robot operating system (ROS).
Experimental results of the method are given in Table Images have been captured at different time period
– II and performance of the method has been to create database for this application. So, there will be
illustrated as a detector from Fig. 9 to Fig. 13, and small amount of image shift. It will be considered as
illustrated as a front-liner from Fig. 14 to Fig. 16. disturbance. This type of disturbance has been
minimized by considering threshold value. Fig.6 gives
Flow chart – I: the database images of S point – I and S point – II of
environment – I. Fig. 7 gives the surveillance images
at S point – I and S point – II of environment – I in
both normal and abnormal conditions.

a b
Fig. 6 Real environment images for database, a) S point – I of
Environment – I, b) S point – II of Environment – I

For observation, change in the environment has

been indicated by red colour circle with thick arrow in
both Fig.7 (b) and Fig.7 (d).

a b c d
Fig. 7 Surveillance images, a) Surveillance – I of S point – I,
b)Surveillance – II of S point – I, c)Surveillance – II of S point – II,
Fig. 4 DoD method flowchart for normal and abnormal d) Surveillance – II of S point – II

53
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 Change in the environment has been identified
using the proposed method. Fig. 8 illustrates resultant
images of both environments using the DoD method
for both conditions.

a b c
Fig. 9 Images at surveillance point, a) Database – I, b) Database – II
at S point, c) Test surveillance

a b c d
Fig. 8 Processed images, a) Surveillance – I of S point – I,
b)Surveillance – II of S point – I, c)Surveillance – II of S point –
II, d) Surveillance – II of S point – II
a b
Fig. 10 Processed images, a) Difference of DB – I and DB – II, b)
Table – I shows performance of the DoD method in Difference of DB – I and TS
both environments under normal and abnormal
conditions. Database has been created and computed threshold
TABLE – I: SURVEILLANCE RESULTS ON REAL
value at surveillance points. Fig. 11 illustrates day – 1
ENVIRONMENT IMAGES USING THE DoD METHOD
surveillance images under normal and abnormal
S. S. Surveillance 𝑻𝑵𝑫 𝑵𝑺𝑺 Remark
No Point conditions of a surveillance point.
Normal
1. I I 1000 20
Abnormal
2. II 3203
Normal
3. II I 1000 597
Abnormal
4. II 19343 a b c d
Fig. 11 Surveillance images of day – 1, a) Surveillance – I image, b)
DoD method has been applied on image of
Difference image of surveillance – I, c) Surveillance – II image, b)
surveillance – I at S point – I, shown in Fig. 8 (a). Difference image of surveillance – II
Pixel difference has been computed as 20, which is not
greater than threshold value 1000. So, it has been
No change has been observed at surveillance point
marked as normal condition and the same thing can be
during surveillance – I. So, captured image will be
observed in original and resultant image. Threshold
similar as database image with small amount of
value has been selected based on repeated calibration
disturbance due to little shift. Pixel difference has been
under normal condition. DoD method has been applied
computed as 46, which is less than threshold value 60.
on image of surveillance – II at S point – I, shown in
Therefore, this surveillance point has been marked as
Fig. 8 (b). Pixel difference has been computed as 3203,
normal condition. This can be observed in the
which is greater than threshold value 1000. So, it has
processed image Fig 11 (b). Change has been observed
been marked as abnormal condition and the same thing
at surveillance point during surveillance – II. So,
can be observed in original and resultant image.
captured image will not be similar as database image.
Similarly, Fig. 8 (c) and Fig. 8 (d) have been marked
Pixel count has been computed as 206, which is
as normal and abnormal conditions of S point – II for
greater than threshold value 60. Therefore, this
surveillance – I and surveillance – II respectively.
surveillance has been marked as abnormal condition.
After getting satisfied results on real environment
The change has been marked with red circle for
images, Practical experiments have been conducted on
general observation. This can be observed in Fig 11 (c)
the DoD method using quad-copter and RPi 3B+
and (d).
(Raspberry model 3B+) as companion computer for
Surveillance process has been repeated at the same
online process in off board mode. This method has
surveillance point on different days with the fixed
been tested for front view scenario. Databases DB – I
database and fixed threshold value. The following Fig.
and DB – II have been created at surveillance point
12 and Fig. 13illustrate the surveillance images of day
and calculated threshold value with the help of trail
– 2 and day – 3.
run under normal condition. Fig. 9 gives the front view
database images at S point.

54
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 training with large dataset to minimize false detection.
False detection has been illustrated in Fig. 15.

a b c d
Fig. 12 Surveillance images of day – 2, a) Surveillance – I image, b)
Difference image of surveillance – I, c) Surveillance – II image, b) Fig. 15 False detection of abnormal condition using HOG with SVM
Difference image of surveillance – II
This type of false detections can be minimized or
removed using the DoD method as front-liner along
with advance technique. Firstly, DoD method will
detect normal or abnormal condition during
surveillance. If it detects abnormal condition then it
a b c d goes with the HOG – SVM technique for the presence
Fig. 13 Surveillance images of day – 3, a) Surveillance – I image, b) of a person in the surveillance area. This surveillance
Difference image of surveillance – I, c) Surveillance – II image, b) process has followed the flow chart – II as given in
Difference image of surveillance – II
Fig. 5.
DoD method has been used to overcome the above
Summary of the experiment results are given in the
false detection. Fig. 16 illustrates process of the DoD
Table – II.
method. Here, 𝑁𝑆𝑆 value for the Fig. 16 (b) is less
TABLE – II: EXPERIMENTAL RESULTS OF ONLINE than threshold value 𝑇𝑁𝐷 . So, it has been marked as
SURVEILLANCE USING THE DoD METHOD normal condition.
S.No Day Surveillance 𝑻𝑵𝑫 𝑵𝑺𝑺 Remark At every surveillance point, HOG – SVM will be
46 Normal followed once DoD detects the abnormal condition at
1. I I Condition
206 Abnormal
surveillance point. DoD method draws less current
2. II Condition when compared with HOG – SVM. So, this process
52 Normal will help in reducing the power consumption during
3. II I 60 Condition surveillance.
190 Abnormal
4. II Condition
50 Normal
5. III I Condition
1505 Abnormal
6. II Condition
a b
HOG – SVM classifier has been used to detect Fig. 16 Surveillance image and processed image using DoD, a)
Surveillance image, b) Resultant image of surveillance
normal and abnormal condition and also to detect the
presence of a person in surveillance area.

a b c
Fig. 17 Detection of abnormal condition, a) Surveillance image, b)
a b Image using DoD, c) Resultant image using HOG – SVM classifier.
Fig. 14 Surveillance images using HOG with SVM, a) Normal, b)
Abnormal, detected one person
Threshold value has been decided by comparing
collected dataset images at different time period to
HOG – SVM classifier has been detected the
avoid false positive and false negative. Positive false
presence of a person and marked the surveillance point
happens only when there is a significant shift in UAVs
as abnormal. Fig. 14 illustrates performance of the
position (i.e., more than a meter shift in UAVs original
HOG – SVM classifier.
surveillance position) due to poor GPS signals or due
If any structure in the environment matches with
to heavy wind blows while surveillance in a patchy
trained model, it will be detected as abnormal
type of field or a field with more objects with different
condition. This is false detection. Here, Tree structure
sizes. False positive will not happen for simple plane
has been marked as the presence of a person and
type of fields even though there is a shift in UAVs
detected surveillance point as abnormal condition.
position.
This is false detection. So, It has been required more

55
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.
 This paper mainly focused on the design and [8]. S. Sambolek and M. Ivasic-Kos, "Automatic Person
implementation of new method with reduced dataset Detection in Search and Rescue Operations Using Deep CNN
for the application as a detector and as a front-liner. Detectors," in IEEE Access, vol. 9, pp. 37905-37922, 2021.
[9]. F. Yang, S. Sakti, Y. Wu and S. Nakamura, "A Framework
So, detailed study on this method with different
for Knowing Who is Doing What in Aerial Surveillance
camera view angles, comparison with other methods
Videos," in IEEE Access,vol. 7, pp. 93315-93325, 2019.
and power consumption details will be discussed in [10]. I. Lahouli, Z. Chtourou, R. Haelterman, G. De Cubber and R.
future. Attia, "A Fast and Robust Approach for Human Detection in
Thermal Imagery for Surveillance Using UAVs,"2018 15th
V. CONCLUSION International Multi-Conference on Systems, Signals &
Devices (SSD),2018, pp. 184-189.
In this paper, a difference of difference (DoD) [11]. M. Piccardi, "Background subtraction techniques: a review,"
method has been proposed to detect normal and 2004 IEEE International Conference on Systems, Man and
abnormal conditions of a boundary while surveillance. Cybernetics (IEEE Cat. No.04CH37583), 2004, pp. 3099-
The work shows performance of the method as a 3104 vol.4.
detector and as a front-liner along with HOG – SVM [12]. S. S. Mohamed, N. M. Tahir and R. Adnan, "Background
modelling and background subtraction performance for object
classifier for surveillance application.
detection," 2010 6th International Colloquium on Signal
Implementation of the method on real images has Processing & its Applications, 2010, pp. 1-6.
been done and achieved satisfactory results in [13]. S. Brutzer, B. Höferlin and G. Heidemann, "Evaluation of
detecting normal and abnormal conditions of different background subtraction techniques for video
surveillance points. Practical experiments have been surveillance,"CVPR 2011, 2011, pp. 1937-1944.
conducted successfully on this method as a detector for [14]. A. Shahbaz, J. Hariyono, and K. H. Jo, “Evaluation of
three days with fixed database using quad-copter for background subtraction algorithms for video surveillance,” in
online surveillance, and as a front-liner along with 2015 21st Korea-Japan Joint Workshop on Frontiers of
Computer Vision (FCV), Jan 2015, pp.1–4.
HOG – SVM classifier to detect any normal or [15]. Z. Zaheer, A. Usmani, E. Khan and M. A. Qadeer, "Aerial
abnormal condition and also presence of a person at surveillance system using UAV,"2016 Thirteenth
the boundary. International Conference on Wireless and Optical
False detection problem with HOG – SVM has Communications Networks (WOCN), 2016, pp. 1-7.
been overcome by using the DoD method as a front- [16]. D. Avola, G. L. Foresti, N. Martinel, C. Micheloni, D.
liner. Pannone and C. Piciarelli, "Aerial video surveillance system
for small-scale UAV environment monitoring," 2017 14th
REFERENCES IEEE International Conference on Advanced Video and
Signal Based Surveillance (AVSS), 2017, pp. 1-6.
[17]. N. H. Motlagh, M. Bagaa and T. Taleb, "UAV-Based IoT
[1]. Rohan, M. Rabah and S. -H. Kim, "Convolutional Neural
Platform: A Crowd Surveillance Use Case," in IEEE
Network-Based Real-Time Object Detection and Tracking for
Communications Magazine, vol. 55, no. 2, pp. 128-134,
Parrot AR Drone 2," in IEEE Access, vol. 7, pp. 69575-
February 2017.
69584,2019.
[18]. H. Teng, I. Ahmad, A. Msm and K. Chang, "3D Optimal
[2]. R. L. Galvez, A. A. Bandala, E. P. Dadios, R. R. P. Vicerra
Surveillance Trajectory Planning for Multiple UAVs by
and J. M. Z. Maningo, "Object Detection Using
Using Particle Swarm Optimization With Surveillance Area
Convolutional Neural Networks, "TENCON 2018 - 2018
Priority," in IEEE Access, vol. 8, pp. 86316-86327, 2020.
IEEE Region 10 Conference, 2018.
[19]. S. V. A. Kumar, E. Yaghoubi, A. Das, B. S. Harish and H.
[3]. S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN:
Proença, "The P-DESTRE: A Fully Annotated Dataset for
Towards real time object detection with region proposal
Pedestrian Detection, Tracking, and Short/Long-Term Re-
network,” IEEE Trans. Pattern Anal. Mach. Intell., Vol. 39,
Identification From Aerial Devices," in IEEE Transactions
no. 6, pp. 1137 – 1149, Jun. 2017.
on Information Forensics and Security, vol. 16,pp.1696-
[4]. Z. Tong, L. Jieyu and D. Zhiqiang, "UAV Target Detection
1708, 2021.
based on RetinaNet,"2019 Chinese Control And Decision
[20]. J. Park, S. Choi, I. Ahn and J. Kim, "Multiple UAVs-based
Conference (CCDC), 2019, pp. 3342-3346.
Surveillance and Reconnaissance System Utilizing IoT
[5]. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y.
Platform," 2019 International Conference on Electronics,
Fu, et al., "SSD: Single shot MultiBox detector", Proc.
Information, and Communication (ICEIC), 2019, pp. 1-3.
ECCV, pp. 21-37, 2016.
[21]. Marianne Bertrand, Esther Duflo, Sendhil Mullainathan,
[6]. J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You
“How Much Should We Trust Differences-In-Differences
only look once: Unified, real time object detection,” in Proc.
Estimates?,” The Quarterly Journal of Economics, Volume
IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun.
119, Issue 1, February 2004, Pages 249–275.
2016, pp. 779 – 788.
[7]. N. Dalal and B. Triggs, "Histograms of oriented gradients for
human detection", IEEE-CVPR, vol. 1, pp. 886-893, June
2005.

56
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on July 25,2025 at 11:46:37 UTC from IEEE Xplore. Restrictions apply.