2020 6th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

Drone Detection Using Convolutional Neural

Networks

Fatemeh Mahdavi Roozbeh Rajabi

Faculty of Electrical and Computer Engineering Faculty of Electrical and Computer Engineering
Qom University of Technology (QUT) Qom University of Technology (QUT)
Qom, Iran Qom, Iran
mahdavi.f@qut.ac.ir rajabi@qut.ac.ir

Abstract— In image processing, it is essential to detect and Fig. 1 make it challenging to detect, identify, and correctly
track air targets, especially UAVs. In this paper, we detect the localize them ]2[. because of high variability among the
flying drone using a fisheye camera. In the field of diagnosis and objects of a similar kind, object detection in the physical
classification of objects, there are always many problems that environment is an onerous duty, furthermore changes in
prevent the development of rapid and significant progress in appearance, illumination, and perspective reduce the
this area. During the previous decades, a couple of advanced efficiency of a detector. Most of the detectors perform
classification methods such as convolutional neural networks ineffectively within the case of changes to the scale and
and support vector machines have been developed. In this study, distortion]2[.
the drone was detected using three methods of classification of
convolutional neural network (CNN), support vector machine Classifications can be divided into available and
(SVM), and nearest neighbor. The outcomes show that CNN, advanced categories. General classification methods include
SVM, and nearest neighbor have total accuracy of 95%, 88%, Maximum likelihood classification (MLC) and Minimum
and 80%, respectively. Compared with other classifiers with the distance classification (MDC). Advanced classifications,
same experimental conditions, the accuracy of the convolutional artificial neural networks (ANNs), Decision trees, support
neural network classifier is satisfactory. vector machines, and object-oriented classifications can also
be mentioned ]3[ .
Keywords—Drone Detection; Bird Detection; Classification;
Convolutional Neural Network; Support Vector Machine; Nearest In general, the classification is done in several stages in
Neighbor such a way that first the required data, which are called train
data, are collected in various ways. then the selected
I. INTRODUCTION classification algorithm is applied to these images, and the
whole Images are classified Then the accuracy of the desired
With the fast advancement within the field of unmanned classification can be examined using test data ]3[.
vehicles and innovation utilized to build them, the number of
drones produced for military, commercial, or recreational The remains of the paper are as follows: In Section II, it
purposes increments pointedly as time passes. This presents a brief of related work. Section III, shows the
circumstance presents urgent protection and security dangers proposed methods. Analysis of the simulation results is
when cameras or weapons are appended to the drones. Thus, presented in Section IV. Then, some conclusions are
detection of the position and properties, similar to speed and discussed in Section V.
track, of drones before an undesirable happening, has gotten
to be vital]1[.
Speed and move capability of drones, their similarity to
birds in appearance when seen from a distance, as shown in

Fig. 1. Drone Detection [1]

 2020 6th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

II. RELATED WORK In Fig. 2, using the 16-NN rule, the point denoted by a
The author of [4] proposed a drone detection solution black dot is classified into the class of the blue points. Out of
based on the Radio Frequency (RF) emitted during the live the sixteen nearest neighbors, five are red, and eleven are
communication session between the drone and its controller blue. The circle indicates the area within which the sixteen
using a Deep Learning (DL) technique, namely, the nearest neighbors lie]11[ .
Convolutional Neural Network (CNN). Mejias et al. used
morphological pre-preparing and Hidden Markov Model
channels to distinguish and track small unmanned vehicles
]5[.
In [6], the authors presented a comprehensive review of
the literature, addressing the concept of drone detection
through object detection and classification using Machine
Learning technologies. Essentially, the application of machine
learning facilitate the drone detection in a binary classification
model as “drone” or “no drone”. However, some research in
the literature goes beyond the traditional classification to a
multiclass classification which identifies drone types.
Fig. 2. 16- Nearest Neighbor
Goke et al. utilized cascaded boosted classifiers with some
local highlight descriptors ]7[. Using a Doppler radar system, B. Support Vector Machine
Michael et al. [8] detected a flying drone in which data would
be gathered by a two-channel Doppler radar and used for In a variety of situations, we need to categorize the
Range-Doppler-azimuth processing. In [9], the drone is information we have. In machine learning, this is called
detected using a convolutional neural network by audio classification. A support vector machine (SVM) is a set of
signals received by the microphone. Because the drone's monitoring algorithms that analyzes data. In SVM, some
harmonic properties are different from objects such as labeled data must be used to classify information ]12[. In
motorcycles, etc. that make similar noises, it has used the general, data in two-dimensional space is separated by a line
Short-Time Fourier Transform (STFT) to extract features. In in three-dimensional space by a plane, and in n-dimensional
[10], drone is detected using frequency-modulated continuous space by a hyperplane.
wave (FMCW) radar system. The system uses antennas with The Support vector is closest data to the decision
a frequency band of 11 to 11.15 GHz to produce a linear wave boundary, and margin is also the distance of the support
of sawtooth antennas with a gain of 33.7dBi. Metal holes are vectors to the decision boundary. The goal is maximize the
used between and on both sides of the antenna to minimize distance of the support vectors from the decision boundary.
path interference. In this method, the drone is detected at a Relation (1) represent the decision boundary equation that w
distance of more than 500 meters. is slope and b is bias of line. The purpose of determining w
There are several methods for detecting Drones. In this and b is not only to classify the data correctly but also to have
article, machine learning strategies such as, Convolutional the decision distance as far as possible with the support
Neural Network (CNN), K-Nearest Neighbor (KNN), Support vectors which is expressed mathematically in equation (2)
Vector Machine (SVM) are utilized and then compared. and (3) ]3[
wt 𝑥+b=0 (1)
III. METHODS
As mentioned in this article, machine learning strategies
such as, convolutional neural network, support vector +1 ,if xt ∈c1 (2)
machine, and nearest neighbor, have been utilized to detect x=(xt ,yt ) yt = { }
-1 ,if xt ∈c2
drones. In this segment, the presentation of the proposed
methods will be inspected. The characteristics of the layers
used in CNN will also be stated.
wt 𝑥 𝑡 +b ≥+1 ,for yt =+1 (3)
{ t 𝑡 }
w 𝑥 +b ≥-1 ,for yt = -1
A. K-Nearest Neighbor
The k-nearest neighbor is an optimization issue to
discover the nearest dots in metric spaces. The nearest X is a point on the decision boundary, and w is vector
neighbor algorithm is used as the first classification model for determine slope ]13[ . Thus, by finding w and b, the
because this model is simple and is robust against a set of decision boundary equation can be obtained.
noisy data and is also useful and applicable for a large data
set. It is the form that assigns data that does not belong to a
category to the group that has the most data around
anonymous data.
 2020 6th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

C. Convolutional Neural Network IV. RESULTS

Convolutional neural networks (CNN) are used for
analysis in machine learning. This method is very efficient A. Datasets
and is one of the most common ways in various The proposed methods use films in which there are drones
applications of computer vision. and birds. These films are converted into frames by
MATLAB software. After extracting the video frames, the
Neural network architecture is presented in different ones that do not have a drone and a bird are not used, and in
forms depending on the type and number of layers, that the the other frames, we separate the images of the drone and the
most common of which are multilayer perceptron ]3[. The birds and put them in two separate folders, one folder
function of this method is based on Equation (4), in which θ containing the images of drones and the other include images
represents the threshold, 𝑤 𝑡 represents the weight vector of of birds. As a result of this operation, 712 images of drones
the coefficients, and x represents the input vector ]3[. and birds were collected. When choosing a set of images, it
g=f (wt x + θ) (4) should be noted that the images in question should not have
marginal and unimportant goals as much as possible, because
A multilayer perceptron consists of an input layer, a these goals can destroy the performance of the program ]15[ .
hidden layer, and an output layer, each containing a certain Eventually, about 80% of this data is utilized for training
number of neurons. The choice of the number of hidden and 20% for testing. In this way, images of 500 birds and
layers and neurons in each of them is significant because if drones are randomly selected and provided to the algorithm
their number is small, the network will face a lack of learning for training at each execution. The residue of the images are
resources to solve complex problems. if it is high, it will used for testing.
cause Network training time increases ]14[. In total, a CNN
network comprises of three primary layers: the convolution This algorithm is implemented in the MATLAB software
layer, the pooling layer, and the fully connected layer. environment.
The pooling layer reduces the size, also decreases the
number of parameters, and reduces the volume of TABLE I. HOG FEATURE EXTRACTION FOR SAMPLE DRONE AND
BIRD
calculations, which in turn helps to control the fitting
phenomenon. In the j map, the 𝑎𝑖𝑗 attribute can be defined in Image Feature Extraction
the ith layer as follows:
a i =f (wn ∗ a i−1 +bij ) (5)
Where f (. ) is an activation function described as f (x)
= max (0, x), it is called the ReLU layer. w is the core of
the filters, n is the index of the feature map set that
Attached to the feature map in (i-1) layer. ∗ is the
convolution operator, and 𝑏𝑖𝑗 is the bias value of the current
feature map. w and b must be trained to be able to extract
better features .
In each convolutional neural network, there are two
steps to preparing. Feedforward stage and backpropagation
stage. In the Feedforward stage, the input image is fed to
the system. This operation is nothing but multiplying the
point between the input and the parameters of each neuron B. Pre processing
and finally performing convolutional operations in each Due to the different sizes of the images in the bird and
layer. The network output is then calculated. drone data, before executing the main parts of the algorithm,
all the images are resized to the same dimensions (48 × 48)
to process the algorithm better.
1) Proposed CNN Structure
We propose CNN structure as indicated by the
multifaceted nature of our information. According to Fig. 1) Extract feature
3, the system consist of three max-pooling layers, three The purpose of feature extraction is to transform the
convolution layers and two fully connected layers. The image into a set of feature vectors that can describe the image
function applied to the data is soft max. Epoch was also well and be used in the classification and recognition phase.
considered 80. The epochs are repeated until the error
reaches a fixed value and does not change . Table I shows how to extract features from drones and
birds in the database for SVM and KNN methods (the
directional gradient histogram feature is used [16]). CNN can
leave out complex data pre-processing and extract feature
automatically [17].
 2020 6th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

Fig. 3. The proposed CNN structure

To test the results, different criteria have been used. These

Finally, the general process used in this paper is shown in criteria include:
Fig. 4
• Accuracy: The number of samples that the system has
successfully detected.
• Sensitivity: How many of the available correct
samples have been selected.
• Precision: How many of the selected samples are
valid. These evaluation metric are given in formulas 6
- 8.

TP :True positive - accurate positive forecast.

TN :True negative - inaccurate positive forecast.
FP :False positive - accurate negative forecast.
FN :False negative - inaccurate negative forecast.

TP+TN (6)
Accuracy=
TP+FP+FN+TN

TP (7)
Sensitivity=
TP+FN
TP (8)
Precision=
TP+FP
Table II shows the classification test using the test data.
Fig. 4. Flowchart of the proposed drone detection algorithm As it turns out, convolutional neural network classifiers
perform better in drone detection than other classifiers.
C. Operation of Classification
According to Fig. 5.a, in exceptional cases such as the TABLE II. COMPARISON RESULTS OF THE CLASSIFIERS
drone being halfway, since it is very similar to a bird, even
the CNN algorithm is not able to detect the drone and Classifier
Sensitivity Precision Accuracy
misidentifies the bird . Fig. 5.b, because the two birds are )%( )%( )%(
very close to each other, the KNN classifier misdiagnoses
the drone but is correctly identified by CNN and SVM. As CNN 93 97 95
shown in Table II, KNN performs worse than SVM and
CNN . SVM 91 86 88

KNN 94 74 80

a b
Fig. 5. Examples of undetectable samples
 2020 6th Iranian Conference on Signal Processing and Intelligent Systems (ICSPIS)

Table III shows that the more convolutional layers and REFERENCES
Epoch, the more accurate the diagnosis, but the more time is [1] C. Aker and S. Kalkan, "Using deep networks for drone detection," in
needed for training. It should be noted that this increase will 2017 14th IEEE International Conference on Advanced Video and
improve performance to some extent, will not have much Signal Based Surveillance (AVSS), 2017, pp. 1-6.
effect on program performance. [2] M. Saqib, S. D. Khan, N. Sharma, and M. Blumenstein, "A study on
detecting drones using deep convolutional neural networks," in 2017
14th IEEE International Conference on Advanced Video and Signal
TABLE III. COMPARISON RESULTS OF CNN BASED ON THE Based Surveillance (AVSS), 2017, pp. 1-5.
NUMBER OF LAYERS [3] A. Richards J., Remote sensing digital image analysis, fifth ed.
springer, 2013 p. 494
two Conv layer four Conv layer
[4] S. Al-Emadi and F. Al-Senaid, "Drone Detection Approach Based on
CNN Properties Radio-Frequency Using Convolutional Neural Network," 2020 IEEE
60 Epoch 80 Epoch International Conference on Informatics, IoT, and Enabling
Technologies (ICIoT), Doha, Qatar, 2020, pp. 29-34, doi:
10.1109/ICIoT48696.2020.9089489.
Time 9min, 56 sec 14 min, 28sec [5] L. Mejias, S. McNamara, J. Lai, and J. Ford, "Vision-based detection
and tracking of aerial targets for UAV collision avoidance," in 2010
IEEE/RSJ International Conference on Intelligent Robots and Systems,
Accuracy (%) 85 95 2010, pp. 87-92.
[6] B. Taha and A. Shoufan, "Machine Learning-Based Drone Detection
and Classification: State-of-the-Art in Research," in IEEE Access, vol.
7, pp. 138669-138682, 2019, doi: 10.1109/ACCESS.2019.2942944.
The specifications of the system with which the above
[7] G. U. o. F. Gokc ¨ ¸e, E. S¸ ahin, and S. Kalkan. , "Vision-Based
results are obtained are as follows: Detection and Distance Estimation of Micro Unmanned Aerial
Vehicles," 2015.
[8] M. Jian, Z. Lu, and V. C. Chen, "Drone detection and tracking based
• Intel Core i5-4200M (2.5GHZ) on phase-interferometric Doppler radar," in 2018 IEEE Radar
• 2 GB DDR3 L Memory Conference (RadarConf18), 2018, pp. 1146-1149.
• 1TB HDD [9] Y. Seo, B. Jang, and S. Im, "Drone Detection Using Convolutional
Neural Networks with Acoustic STFT Features," in 2018 15th IEEE
For example, Fig. 6 shows a video frame in which the International Conference on Advanced Video and Signal Based
drone is correctly separated from the birds. Surveillance (AVSS), 2018, pp. 1-6.
[10] Park and S. Park, "Configuration of an X-band FMCW radar targeted
for drone detection," in 2017 International Symposium on Antennas
and Propagation (ISAP), 2017, pp. 1-2.
[11] S. Theodoridis and K. Koutroumbas, "Pattern recognition," Academic
Press, 2009.
[12] R. Fernández, Marcos, José, "Computer vision for Pedestrian detection
using Histograms of Oriented Gradients," 2014.
[13] A. Keshavarz, Qasemian Yazdi,'' A fast vector machine-based backup
algorithm for classifying hyperspectral images using spatial
correlation'', Electrical Engineering and Computer Engineering Iran, in
Fig. 6. Drone Detection Using CNN Persian
[14] M. A. a. P. L. Bartlett, Neural network learning: Theoretical
foundations. Cambridge university press, 2009.
V. CONCLUSION [15] N. Dalal, "Finding People in Images and Videos," 07/17 2006.
Today, drones are used for various purposes. These [16] K. Mihreteab, M. Iwahashi, and M. Yamamoto, "Crow birds detection
include goals such as protection, passage to small areas, using HOG and CS-LBP," in 2012 International Symposium on
Intelligent Signal Processing and Communications Systems, 2012, pp.
espionage uses, and so on. The small size and low noise of 406-409.
the drone make its role more critical. Therefore, drone
[17] X. Li, Z. Huang, F. Wang, X. Wang, and T. Liu, "Toward
detection has also become important for security issues. In Convolutional Neural Networks on Pulse Repetition Interval
this paper, the proposed algorithm has two stages of training Modulation Recognition," IEEE Communications Letters, vol. 22, no.
and testing. In the training phase, the properties of all the 11, pp. 2286-2289, 2018.
images used were extracted, and a convolutional neural
network was trained using these properties. Then, in the next
step of experiment, the input image was classified by a
trained system. Therefore, the confusion between some
similar parts of the drone and the bird was eliminated using
the convolutional neural network classification model. The
convolutional neural network model improves the efficiency
in comparison of the support vector machine and the nearest
neighbor in detecting similar parts of the two objects.