Content uploaded by Eduardo Kugler Viegas
Author content
All content in this area was uploaded by Eduardo Kugler Viegas on Aug 31, 2023
Content may be subject to copyright.
Benchmarking Audio-based Deep Learning Models
for Detection and Identification of Unmanned Aerial
Vehicles
Sai Srinadhu Katta, Sivaprasad Nandyala, Eduardo Kugler Viegas, Abdelrahman AlMahmoud
Secure Systems Research Centre (SSRC)
Technology Innovation Institute (TII)
Abu Dhabi, U.A.E.
{sai, siva, eduardo, abdelrahman}@ssrc.tii.ae
Abstract—Over the last few years, Unmanned Aerial Vehicles
(UAVs) have become increasingly popular for both commercial
and personal applications. As a result, security concerns in both
physical and cyber domains have been raised, as a malicious
UAV can be used for the jamming of nearby targets or even for
carrying explosive assets. UAV detection and identification is a
very important task for safety and security. In this regard, several
techniques have been proposed for the detection and identification
of UAVs, in general, through image, audio, radar, and RF based
approaches. In this paper, we benchmark the detection and
identification of UAVs via audio data from [1]. We benchmarked
with widely used deep learning algorithms such as Deep Neural
Networks (DNN), Convolutional Neural Network (CNN), Long
Short Term Memory (LSTM), Convolutional Long Short Term
Memory (CLSTM) and Transformer Encoders (TE). In addition to
the dataset of [1], we collected our own diverse identification audio
dataset and experimented with Deep Neural Networks (DNN).
In a UAV detection task, our best model (LSTM) outperformed
the best model of [1] (CRNN) by over 4% in accuracy, 2% in
precision, 4% in recall and 4% in F1-score. In UAV identification
task, our best model (LSTM) outperformed the best model of [1]
(CNN) by over 5% in accuracy, 2% in precision, 4% in recall
and 3% in F1-score.
Index Terms—Audio, Deep learning, Detection, Identification,
UAVs
I. INTRODUCTION
The popularity of Unmanned Aerial Vehicles (UAV), com-
monly referred as drones, have significantly increased over the
last few years, followed by the technological advancements
of their on-board components. In practice, modern UAVs
have enabled the deployment of user customized solutions,
which are able to analyze the data from a variety of UAV
hardware components including the camera, microphones,
LiDAR, accelerometer, GPS among others [2]. This ease of
customized UAV solutions has paved the way for several
autonomous UAV applications, such as object delivery, field
surveillance, and even border control etc. [3].
Unfortunately, modern UAVs can also be used for malicious
purposes. Unsurprisingly, drone-based attacks can have a
significant negative impact on the economy, safety and security.
As a result, in recent years several works have been proposed
for the detection and identification of nearby UAVs [4]–[8].
Detection and identification of nearby UAVs are generally
achieved using vision-based analysis [9], radio fingerprint
detection [10], radar-based identification [11], and microphone-
based approaches [12], [13]. In this work we focus on
UAV detection and identification by audio-based approaches.
Machine Learning (ML) based techniques are predominantly
used for analyzing audio acoustic UAV signatures for UAV
detection and identification [12]–[17].
We benchmarked UAV drone detection and identification
from publicly available audio dataset [1], via commonly
used deep learning algorithms namely DNN, CNN, LSTM,
CLSTM and TE. In addition to [1], we collected our diverse
identification audio dataset and experimented with DNN models
on it.
The main contributions of this paper are as follows:
•
We benchmarked various deep learning models namely
DNN, CNN, LSTM, CLSTM and TE on publicly available
UAV detection and identification audio dataset [1]. Our
models perform significantly better than the [1].
•
In addition to the dataset of [1], we collected our own
diverse identification audio dataset consisting of 7 different
categories of UAVs namely no-UAV, drone, helicopter,
drone-membo, drone-bebop, airplane, and drone-hovering.
We experimented with DNN models on this and results
are promising.
The remainder of the paper is organized as follows. Section
II describes the literature review. Section III presents the audio-
based scheme for UAV detection and identification through
deep learning architectures. Section IV talks about experimental
results and discussion. Finally, Section V concludes our work.
II. LITERATURE REVIEW
Over the last years, several techniques have been proposed
for UAV detection and identification, ranging from video [5],
radio frequency [6], [18], thermal imaging [7], radar [11],
and more recently audio-based [12], [14], [19]–[22]. In such
a context, several approaches have been proposed for audio-
based UAV detection and identification techniques, as they can
be easily deployed due to their negligible equipment costs, and
the promising accuracy results reported in the literature.
7
2022 Workshop on Benchmarking Cyber-Physical Systems and Internet of Things (CPS-IoTBench)
978-1-6654-7038-4/22/$31.00 ©2022 IEEE
DOI 10.1109/CPS-IoTBench56135.2022.00008
2022 Workshop on Benchmarking Cyber-Physical Systems and Internet of Things (CPS-IoTBench) | 978-1-6654-7038-4/22/$31.00 ©2022 IEEE | DOI: 10.1109/CPS-IoTBench56135.2022.00008
Authorized licensed use limited to: Pontifica Universidade Catolica do Parana. Downloaded on August 31,2023 at 17:45:45 UTC from IEEE Xplore. Restrictions apply.
To achieve such a goal, UAV audio-based detection and
identification is often implemented through four sequential
modules. First, the Data Acquisition module collects sound
samples from a given microphone. In general, the collected data
is evaluated according to a given predefined time interval, e.g.
every
1
second. Then, the Feature Extraction module extracts
a set of behavioral features from the analyzed audio sample,
compounding a feature set. Several techniques can be used
to fulfill such a task in audio-based detection, including the
extraction of the audio spectrogram [17], [23], and the building
of the audio coefficients [8]. Finally, the Detection/Identification
module classifies the built feature set in one of the selected
classes.
In recent years, a plethora of highly accurate audio-based
detection and identification approaches have been proposed
for the detection and identification tasks in UAV detection
and identification [1], [14], [18], [20], [24]. In general, authors
resort to ML approaches, typically implemented through pattern
recognition techniques. To fulfill such a task, the operator
typically relies in a two-phase process, namely training, and
testing.
The training step aims at training of the ML model with
the training dataset and choosing the best model on the
validation dataset. The testing phase evaluates the final model
detection and identification performance metrics. In practice,
the performance measurements obtained at testing phase are
expected to be evidenced when the designed system is deployed
in production environments.
III. PROPOSED APPROACH
The tasks here are detection and identification of UAV via
UAV audio signatures. The overall procedure of our method is
shown in Figure 1.
Almost all the related existing literature works [12], [13],
[15]–[17] exclusively followed the representation learning
techniques, we have also picked that as a design choice for
the implementation. Previous results on various settings of this
task gave state-of-the-art results. The audio data is sent through
feature extraction, the extracted features are passed through the
deep learning models giving the identification and detection
results. The end-to-end model is as follows: first an audio is sent
through mel frequency cepstral coefficients (MFCC) feature
extraction,followed by a deep learning model of choice (DNN,
CNN, LSTM, CLSTM, TE) giving us the result of detection
and identification. Based on the result safety/alert actions can be
performed accordingly. The detailed deep learning architectures
are as follows:
1) Deep Neural Network (DNN): The DNN is made of
fully-connected layers and non-linear activations. The input
to the DNN is the flattened MFCC features, which feeds into
a stack of hidden fully-connected layers. At the output is a
linear layer followed by a softmax layer generating the output
probabilities of the classes.
2) Convolutional Neural Network (CNN): CNNs exploit
the local temporal and spectral correlation in the features via
2D convolution. The input to the CNN is the MFCC features,
Fig. 1. Overview of the proposed deep learning model for audio-based
detection and identification of UAVs.
which feeds into a stack of convolution layers. At the output
is a linear layer followed by a softmax layer generating the
output probabilities of the classes.
3) Long Short-Term Memory (LSTM): LSTMs are known
to model long term dependencies and are shown to work very
well on various sequence modelling tasks. The input to the
LSTM is the MFCC features, and the whole flattened output
sequence is fed to a linear layer followed by softmax for output
probabilities of the classes.
4) Convolutional Long Short-Term Memory (CLSTM):
CLSTMs are combination of convoultion followed by LSTMs.
This has the benefits of both CNNs and LSTMs. CLSTMs
exploit the local temporal and spectral correlation, and model
the long term dependencies well. The input to the CLSTM is
the MFCC features, which has convolution followed by LSTM,
and the whole flattened output sequence is fed to a linear layer
followed by softmax for output probabilities of the classes
5) Transformer Encoder (TE): Transformers are shown to be
the fundamental block for SOTA on various sequence modelling
tasks across various domains. In this work we use only the
Transformer Encoder part. The input is the MFCC features,
and the whole output sequence is fed to a linear layer followed
by softmax for output probabilities of the classes.
IV. EXPERIMENTAL RESULTS AND DISCUSSION
Our proposed scheme was evaluated and compared consid-
ering both literature and our own built dataset. The evaluation
aims at answering the following research questions, as follows:
(RQ1)What is the detection accuracy of audio-based techniques
in a publicly available dataset? (RQ2)What is the identification
accuracy when different UAV types are considered? (RQ3)What
is the identification performance impact when more diverse
flying devices are considered?.
The proposed scheme for an audio-based detection and
identification of nearby UAVs is implemented through a
pattern recognition pipeline. Therefore, we consider a given
microphone deployed in a monitored environment (Fig. 1,
Deployed Microphone), which will be used for the collection
and periodic sending of the environment audio samples. For
instance, collection of
1
-sec batches of audio in a predefined
format. The main assumption is that nearby flying UAVs will
8
Authorized licensed use limited to: Pontifica Universidade Catolica do Parana. Downloaded on August 31,2023 at 17:45:45 UTC from IEEE Xplore. Restrictions apply.
produce audio noises that will be captured by the microphone
for further analysis. The collected audio sample is analyzed by
the feature extraction module, which goal is to extract a set
of UAV related features. To achieve such a goal the module
applies an audio filtering technique ( e.g., filtering UAV-related
audio signals through the Mel-Frequency Cepstral Coefficients
(MFCC)), before using it for the detection and identification
tasks. As a result, a portion of non UAV-related audio can
be removed from the analyzed sample, improving the system
generalization even in highly noise environments. Finally, the
extracted filtered feature vector (Fig. 1, Filtered Feature Vector)
is used as input by a deep learning classifier, which outputs
a corresponding event class. In such a case, the used deep
learning model can output the analyzed event label in a two-
class setting, e.g. normal or UAV, or even in a multiclass setting,
e.g. outputting the type of UAV that generated the analyzed
audio.
The next subsections further describes the built datasets and
the performed experiments.
A. UAV Audio Datasets
In general, to properly evaluate ML-based techniques for
audio detection and identification, it is necessary the provision
of huge amounts of labeled data. Unfortunately, due to privacy
issues, only few datasets are publicly available.
In light of this, our work makes use of Abdulla Al-Ali et.
al. [1] dataset which provides more than
1300
audio clips of
drone sounds. We used the dataset mainly for two tasks, one
is detection and other is identification. In detection, the classes
are Drone and Not a Drone. Number of drone samples are
1332 and not a drone samples are 10372. In classification task
the classes are membo, bebop, and not a Drone, the number
of samples are 666, 666, and 10372 respectively. Every file is
of 1sec duration. The data set is split on file level basis. The
drone data is a good representation of the real-world drone
audio.
To increase the model performance, the audio dataset is also
augmented through the introduction of noise data to ensure that
the system will be able to detect and identify the drone’s sound
from similar noises in an environment. From [1], the SNR
levels are not available. For the dataset publicly available, it
was collected in a quite indoor environment with drone flying
and hovering.
Apart from the publicly available dataset, we have also
collected a new dataset with additional diverse flying devices
audio sounds from multiple open sources to evaluate the impact
of model identification performance metrics. The dataset was
built, with
7
UAV types, including no-UAV,drone,helicopter,
drone-membo,drone-bebop,airplane, and drone-hovering. The
majority of audio files for each UAV type was collected for 5
minutes.
For both selected datasets the audio data is made with a
sampling rate of
48
kHz and a linear encoding with
16
bits for
sample. Each input sound window is further segmented into
sub-frames of
20
ms using a moving Hamming window with
overlap of
10
ms. The sub-frames are processed by a bank of
filters to compute the short term feature in both temporal and
frequency domains [20]. For the deep learning algorithms, If
the majority of the frames is labelled with the tag ”drone” into
the given audio segment, it is assumed to recognize a flying
drone in the surrounding environment.
Both datasets are split into
80
% for training,
10
% for
validation and
10
% for testing. The audio files are broken
into chunks of 1second.
B. Model Building
For the Feature Extraction module, for each audio frame,
we compute the Mel-Frequency Cepstral Coefficients (MFCC)
that are commonly used in audio analysis. We make use of
Mel-scale in audio analysis to capture the comparatively higher
energy in lower frequencies compared to higher frequencies in
the range for the compounding of
40
MFCC features, which
are used as input by the selected deep learning model (Fig. 1,
Filtered Feature Vector).
We evaluate
5
commonly used deep learning algorithms for
the audio-based classification task (Fig. 1, Deep Learning),
namely Deep Neural Network (DNN), Convolutional Neural
Network (CNN), Long Short-Term Memory (LSTM), Convolu-
tional Long Short-Term Memory (CLSTM), and Transformer
Encoder (TE).
The DNN was implemented with
3
hidden layers, each with
256
units, coped with a relu activation function, while the
output layer relies in a softmax activation function. The CNN
was implemented with
3
convolutional layers, each with a relu
activation function, a kernel size of
3X3
, and a
1X1
stride,
followed by a hidden layer with a softmax activation function.
There are no pooling layers used in the CNN we implemented,
the input dimension size to the CNN is [1X20X33], given the
small size there was no need to use the pool layers. The CNN
model architecture is not provided in [1], due to this, we are
not able to do a proper comparison in terms of architecture,
we made our design choice to keep the model as smaller as
possible.
The LSTM was implemented with one LSTM layer with
128
units, followed by a hidden output layer with a softmax
activation function. Similarly, the CLSTM makes use of a
convolutional layer with a relu activation function, a kernel
size of
3X3
, and a
1X1
stride, followed by a LSTM layer with
128
units, followed by a hidden output layer with a softmax
activation function. Finally, the TE was implemented with a
positional encoder, number of attention heads as
2
, followed
by a encoder layer with
128
units, and a output layer with a
softmax activation function,
For the model building procedure, for all selected deep
learning algorithms
100
epochs are executed with a batch
size of
128
. As for number of epochs, 100 is similar to
the related works. It is important to note that the used set
of parameters, were set similarly to related works, and no
significant differences were found while varying them. There
is 80% train data, 10% val data, 10% test data. In each epoch
train and val data are used, model is trained on train data and
9
Authorized licensed use limited to: Pontifica Universidade Catolica do Parana. Downloaded on August 31,2023 at 17:45:45 UTC from IEEE Xplore. Restrictions apply.
TABLE I
DETECTION RESULTS OF THE DEEP LEARNING ALGORITHMS AS MEASURED AT THE PUBLICLY AVAILABLE DATASET [1]
Deep Learning Algorithm Accuracy (%) Precision Recall F1-Score
Recurrent Neural Network (RNN [1] 75.00 0.7592 0.6801 0.6838
Convolutional Neural Network (CNN) [1] 96.38 0.9624 0.9560 0.9590
Convolutional Recurrent Neural Network(CRNN) [1] 94.72 0.9502 0.9308 0.9393
Deep Neural Network (DNN) 98.35 0.9661 0.9549 0.9604
Convolutional Neural Network (CNN) 98.85 0.9753 0.9696 0.9724
Long Short-Term Memory (LSTM) 98.93 0.9759 0.9731 0.9745
Convolutional Long Short-Term Memory (CLSTM) 97.78 0.9460 0.9486 0.9473
Transformer Encoder (TE) 98.35 0.9634 0.9489 0.9606
TABLE II
IDENTIFICATION RESULTS OF THE DEEP LEARNING ALGORITHMS AS MEASURED AT THE PUBLICLY AVAILABLE DATASET [1]
Deep Learning Algorithm Accuracy (%) Precision Recall F1-Score
Recurrent Neural Network (RNN [1] 57.16 0.5964 0.5716 0.5562
Neural Network (CNN) [1] 92.94 0.9275 0.9263 0.9263
Convolutional Recurrent Neural Network (CRNN) [1] 92.22 0.9254 0.9223 0.9225
Deep Neural Network (DNN) 98.52 0.9589 0.9439 0.9508
Convolutional Neural Network (CNN) 98.60 0.9553 0.9447 0.9510
Long Short-Term Memory (LSTM) 98.60 0.9480 0.9603 0.9540
Convolutional Long Short-Term Memory (CLSTM) 98.11 0.9457 0.9188 0.9314
Transformer Encoder (TE) 98.19 0.9405 0.9369 0.9386
best model on val data is updated. Finally the best val model
is tested on the unseen test data.
C. Evaluation
The selected deep learning algorithms were evaluated with
respect to their accuracy, precision, recall and F1 scores. To
achieve such a goal, the following classification performance
metrics were used:
•
True-Positive (TP): number of UAV-related audio samples
correctly classified as UAV-related.
•
True-Negative (TN): number of normal samples correctly
classified as normal.
•
False-Positive (FP): number of normal samples incorrectly
classified as UAV-related.
•
False-Negative (FN): number of UAV-related audio sam-
ples samples incorrectly classified as normal.
The F1 score was computed as the harmonic mean of
precision and recall values while considering UAV-related as
positive samples and normal as negative samples, as shown in
Eq. 3.
P recision =TP
TP +FP (1)
Recall =TP
TP +FN (2)
F1=2∗P recision ∗Recall
P recision +Recall (3)
From the literature [1], there were no experiments conducted
on the publicly available dataset with DNN, LSTM and
Transformer techniques. We performed them for a proper
TABLE III
INDIVIDUAL ACCURACIES IN OUR OWN BUILT DATASET.
UAV type Accuracy (%) Prec. Rec. F1-Score
no-UAV 85.71 1.00 0.86 0.92
drone 100 1.00 1.00 1.00
helicopter 66.66 0.67 0.67 0.67
drone-membo 100 1.00 1.00 1.00
drone-bebop 100 1.00 1.00 1.00
airplane 100 1.00 1.00 1.00
drone-hovering 100 0.88 1.00 0.93
OVERALL Dataset 95 0.96 0.95 0.95
comparison to asses the overall improvement for the detection
and classification tasks.
The first experiment, aims at answering RQ1 and evaluates
the identification accuracy of the selected deep learning algo-
rithms over the publicly available dataset [1]. The evaluation
goal is to measure how the selected techniques perform when
using publicly available dataset for audio-based detection of
UAVs. In Table I shows, modeling results of [1] and our own
benchmark results for drone detection task. All our models
outperformed the models in [1] by 4% to more than 20% in
accuracy. Our best model (LSTM) outperformed best model of
[1] (CRNN) by over 4% in accuracy, 2% in precision, 4% in
recall and 4% in F1-score. Our best detection model, LSTM
classifier achieved the highest accuracy of
98.93
%, precision
of 0.9759, recall of 0.9731 with an F1-Score of 0.9745.
The second experiment, aims at answering RQ2, and eval-
uates the UAV identification accuracy of the selected deep
learning algorithms when different UAV types are considered.
To achieve such a goal, the publicly available dataset [1] is
10
Authorized licensed use limited to: Pontifica Universidade Catolica do Parana. Downloaded on August 31,2023 at 17:45:45 UTC from IEEE Xplore. Restrictions apply.
evaluated with different types of UAVs. In Table II shows,
modeling results of [1] and our own benchmark results for
drone identification task. All our models outperformed the
models in [1] by 6% to more than 30% in accuracy. Our best
model (LSTM) outperformed best model of [1] (CNN) by
over 5% in accuracy, 2% in precision, 4% in recall and 3%
in F1-score. Our best identification model, LSTM classifier
achieved the highest accuracy of
98.60
%, precision of
0.9480
,
recall of
0.9603
with an F1-Score of
0.9540
. For instance, our
identification LSTM model was able to achieve
98.60
%of
accuracy, a decrease of only
0.33
% when compared to the
detection scenario (Table I vs. Table II).
The DNN model in table 3 is trained and tested on the newly
collected 7 class UAV dataset. The third experiment aims at
answering RQ3, and evaluates the identification accuracy in
our own built dataset. Table III shows that the classification
accuracy of the DNN classifier in our own built dataset. The
DNN classifier achieved a high accuracy of
95.20
%. Compared
to its counterpart, trained in the publicly available dataset, the
DNN decreased the accuracy by
3.32
%, a marginal decrease
considering that the classifiers is being used in a different and
difficult setting.
V. C ONCLUSIONS
In this work, we benchmarked the detection and identification
of UAVs via audio, through multiple deep learning algorithms
namely DNN, CNN, LSTM, CLSTM, and TE. In addition to
dataset of [1], we also collected our own identification dataset
and built DNN model over it. We have demonstrated that all
of the DNN, CNN, LSTM, CLSTM and TE algorithms are
able to provide significantly higher performance metrics in
comparison to [1]. In detection task, our best model (LSTM)
outperformed the best model of [1] (CRNN) by over 4% in
accuracy, 2% in precision, 4% in recall and 4% in F1-score.
In identification task, our best model (LSTM) outperformed
the best model of [1] (CNN) by over 5% in accuracy, 2% in
precision, 4% in recall and 3% in F1-score.
REFERENCES
[1]
Abdulla Al-Ali et. al. Sara Al-Emadi, “Drone audio dataset,”
https://github.com/saraalemadi /DroneAudioDataset, 2018.
[2]
Eleni Diamantidou Samaras, Stamatios and Antonios Lalas et al., “Deep
learning on multi sensor data for counter uav applications—a systematic
review.,” Sensors 19, vol. 22, pp. 4387, 2019.
[3]
Austin Reg, “Unmanned aircraft systems: Uavs design, development and
deployment,” John Wiley & Sons, pp. 1–365, 2011.
[4]
Beomhui Jang Jangwon Jung Seo Yoojeong and Sung bin Im, “Uav
detection using the cepstral feature with logistic regression,” Sensors,
pp. 219–2222, 2018.
[5]
Changwen Zheng Peng Junkai and Si Lingyu et.al., “Using images
rendered by pbrt to train faster r-cnn for uav detection,” Computer
Science Research Notes CSRN 2802, pp. 770–778, 2018.
[6]
Faith Erden Martins Ezuma and Ismail Guvenc et.al., “Micro-uav
detection and classification from rf fingerprints using machine learning
techniques,” 2019 IEEE Aerospace Conference, pp. 1–13, 2019.
[7]
Emmanuel Zenou Eren Unlu and Nicolas Riviere, “Using shape
descriptors for uav detection,” Electronic Imaging, vol. 1-5, 2018.
[8]
Amin Ullah Jamil Sonain, MuhibUr Rahman and Seyed Sajad Mirjavadi
et.al., “Malicious uav detection using integrated audio and visual features
for public safety applications,” Sensors 20, p. 3923, 2020.
[9]
Liu Hao and Yunfang Ren et al., “Drone detection based on an audio-
assisted camera array,” 2017 IEEE Third International Conference on
Multimedia Big Data(BigMM), p. 402–406, 2017.
[10]
Erden F. Ezuma M. and I. et.al. Guvenc, “Micro-uav detection and
classification from rf fingerprints using machine learning techniques,”
2019 IEEE Aerospace Conference, pp. 1–13, 2019.
[11]
Justin Matheson Philip Church, Christopher Grebe and Brett Owens,
“Aerial and surface security applications using lidar. in laser radar
technology and applications,” International Society for Optics and
Photonics, vol. 10636, pp. 27–38, 2018.
[12]
Chaoqun Yang Chang Xianyu and Zhiguo Shi. et. al., “A surveillance
system for drone localization and tracking using acoustic arrays,” IEEE
10th Sensor Array and Multichannel Signal Processing Workshop (SAM),
pp. 573–577, 2018.
[13]
Ald rich Martinez-Carranza Cabrera-Ponce and Caleb Rascon, “Detection
of nearby uavs using a multi-microphone array on board a uav,”
International Journal of Micro Air Vehicles, vol. 12, pp. 1–10, 2020.
[14]
Abdulla Al-Ali Al-Emadi, Sara and Abdulaziz Al-Ali, “Audio-based
drone detection and identification using deep learning techniques with
dataset enhancement through generative adversarial networks,” Sensors
21, p. 4953, 2021.
[15]
Wansong Zhu Chen Liu and Minghui Zheng, “Audio-based fault diagnosis
method for quadrotors using convolutional neural network and transfer
learning,” 2020 American Control Conference, pp. 1367–1372, 2020.
[16]
Sara Abdulla Al-Ali Amr Mohammad Al-Emadi and Abdul aziz Al-Ali,
“Audio based drone detection and identification using deep learning,”
15th International Wireless Communications and Mobile Computing
Conference (IWCMC), pp. 459–464, 2019.
[17]
Vemula Hari Charan., “Multiple drone detection and acoustic scene
classification with deep learning,” PhD diss., Wright State University,
pp. 1–48, 2018.
[18]
H. Truong P. Nguyen and T. Vu.Matthan et.al., “Drone presence detection
by identifying physical signatures in the drone’s rf communication,” in
Proceedings of the 15th Annual International Conference on Mobile
Systems, Applications, and Services, ACM, vol. 5, pp. 211–224, 2017.
[19]
J Shin S Jeon and H. Yang et.al., “Empirical study of drone sound
detection in real-life environment with deep neural networks,” 2017 25th
European Signal Processing Conference (EUSIPCO), 2017.
[20]
Emiliano Pallotti Bernardini Andrea, Federica Mangiatordi and Licia
Capodiferro, “Drone detection by acoustic signature identification,”
Electronic Imaging, pp. 60–64, 2017.
[21]
Kolamunna Harini and thilini et al., “Droneprint: Acoustic signatures
for open-set drone detection and identification with online data,” Pro-
ceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous
Technologies, pp. 1–31, 2021.
[22]
Darren Haddad Alexander Sedunov and Alexander Yakubovskiy et.al.,
“Stevens drone detection acoustic system and experiments in acoustics
uav tracking,” 2019 IEEE International Symposium on Technologies for
Homeland Security (HST), pp. 1–7, 2019.
[23]
C Park J Kim and JC Gallagher et.al., “Real-time uav sound detection
and analysis system,” 2017 IEEE Sensors Applications Symposium (SAS),
pp. 1–5, 2017.
[24]
Amr Mohammad Al-Emadi, Sara Abdulla Al-Ali and Abdulaziz Al-
Ali, “Audio based drone detection and identification using deep learning,”
2019 15th International Wireless Communications and Mobile Computing
Conference (IWCMC), pp. 459–464, 2019.
11
Authorized licensed use limited to: Pontifica Universidade Catolica do Parana. Downloaded on August 31,2023 at 17:45:45 UTC from IEEE Xplore. Restrictions apply.