Page 52 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 5 – Internet of Everything

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 5

Table 1 – Related work on detection and classi ication of drones using ML techniques.
Litera‑ Source type Features Data process Classi i‑ # of Accuracy Noise
ture method cation UAVs conside-
ration

[10] Drone Slope, kurtosis Several ML algorithms X N/A 96.36% X
RF signals skewness
Drone Channel state
[11] CSI data X N/A 86.6% X
RF signal information
Acoustic
[12] MFCC and LPCC SVM X N/A 96.7% X
waves
Acoustic
[13] STFT features CNN X N/A 99.87% X
waves
CNN for moving body
Camera
[14] RGB arrays detection and kNN for X N/A 93% X
images
detection
Camera CNN on ZF and VGG16
[15] RGB arrays X N/A 0.66 mAP X
images and Fast R‑CNN
2‑D complex‑log‑
[16] Radar signals Spectrogram X N/A 3.27% EER X
Fourier transform
Range Doppler
[17] Radar signals SVM X N/A 98% X
Matrix
Micro‑Doppler PCA feature extraction
[18] Radar signals X 3 94.7% X
signature on spectrograms
Micro‑Doppler
[19] Radar signals CNN and LSTM‑RNN X 5 97.7% X
spectrogram
Micro‑Doppler
[20] Radar signals CNN X 6 94.7% X
signature
Micro‑Doppler
[21] Radar signals signatures SVM X 11 >95% X
through EMD
Micro‑Doppler
[22] Radar signals SVM X 11 95.4% X
signatures
Range Doppler 99.5% and
[23] Radar signals CNN X N/A X
spectrum 54.2% for 0 dB
Statistical features
Drone 88‑94% in
[24] e.g., mean, median, Logistic regression X 8 X
RF signals 0.35 s
RMS
Micro‑Doppler
[25] Radar signals ANN on MLP X 4 Various X
signature
98.13% and
Controller Shape factor,
[26] Several ML algorithms X 17 40% for 0 dB X
RF signals kurtosis, variance
SNR
Time‑series
This Controller signal and CNN X 99.7% and X
work RF signals spectrogram 15 99.5% for
RGB arrays 0 dB SNR

• For the classi ication tasks that involve RF in‑ • In this work, we apply denoising on the spectrogram
gerprinting, variations in the Signal‑to‑Noise Ra‑ images to further improve the performance at
tio (SNR) of the received RF signals is a challenging low SNRs. We tune the spectral density level that
problem. In this work, we also address this prac‑ will appear on the spectrogram image and
tical problem by considering a range of SNR levels ilter out spectral densities lower than the tuned
from −10 dB to 30 dB while training the CNN models. level. Our proposed classifier highly outperforms
Noisy training data is generated by adding arti icial previously published work, especially at low SNRs.
white noise to the original data. When using spec‑
trogram images to train the CNN models, we only
focus on the frequency range of interest, which im‑
proves ication accuracy icantly in com‑
parison with time‑series images.

40 © International Telecommunication Union, 2021