Experimental UAV detection using 4G-LTE-based

passive Radar
Abigael Taylor, Dominique Poullin

To cite this version:

Abigael Taylor, Dominique Poullin. Experimental UAV detection using 4G-LTE-based passive Radar.
IEEE Radar Conference 2023, Nov 2023, SYDNEY, Australia. �hal-04372850�

HAL Id: hal-04372850

https://hal.science/hal-04372850
Submitted on 4 Jan 2024

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Experimental UAV detection using 4G-LTE-based
passive Radar
Abigael Taylor Dominique Poullin
Department of Electromagnetism and Radar Department of Electromagnetism and Radar
ONERA, the French aerospace Laboratory ONERA, the French aerospace Laboratory
Université Paris, Saclay, F-91123 Palaiseau France Université Paris, Saclay, F-91123 Palaiseau France
[email protected] [email protected]

Abstract—This paper presents detection results of drone in a a drone evolving above the surveillance antenna, making the
passive radar framework, that exploits 4G Long Term Evolution environment friendlier that the one that we will consider here.
signals. The results are based on an experiment using a non- In [6], cars were succesfully detected. In [7], a cooperative
cooperative transmitter. The processing of the data is described,
in particular how general passive algorithm can be adjusted to 5G-LTE transmitter was used for a succesful drone detection.
the processing of LTE signals. The results are then discussed. In this paper, drone detection is investigated in a non-
Index Terms—UAV, Passive Radar, LTE, Detection, COFDM, cooperative environment. The main contributions can be sum-
4G. marized as follows:
• a suggestion of a way of processing the data to handle
I. I NTRODUCTION
most of LTE issues that we encountered;
Passive radar exploits non-cooperative external transmitters • an analysis of the received signals;
to achieve the usual tasks one can expect from a radar • presentation and discussion on detection results.
system. Many examples of successful application of passive
This paper is organized as follows. To begin, basic informa-
radar using Digital Video Broadcasting (DVB) or Digital
tion on COFDM and LTE signals will be reminded. The signal
Audio Broadcasting (DAB) can found in the literature [1]. In
processing used will then be introduced. The measurement
particular, passive radar represents an interesting candidate for
campaign will then be briefly described. Afterwards, we will
Unmanned Aerial Vehicle (UAV) detection [2], as it does not
study eigenvalues of the measured signal covariance matrix.
interfere with the surrounding electromagnetic environment,
This will give some enlightening information on the data, and
and can thus be deployed next to sensitive places such as
with what care it should be processed. Finally, results on the
airports. Various illuminators of opportunity and waveforms
detection of the UAV will be presented and commented.
have been studied to respond to the threat represented by
UAVs, such as DVB-Terrestrial [3] and DVB-Satellite [4], II. S IGNAL MODEL
which both offer wide coverage, a valuable quality for UAV
detection. For the last decade, the use of Long-Term Evolution The LTE signals used for our experiment are based on
signals has also been investigated [2], [5]–[7] as it represents COFDM modulation. We will thus proceed to recall the gen-
a widely used wireless communication strategy today, and eral expression of this modulation, and then the specificities
thus can be of interest in particular for urban areas. All these of LTE signals that are of importance for this work will be
signals (DAB, DVB and LTE) are based on Coded Orthogonal presented.
Frequency Division Multiplexing (COFDM) modulation, but
can differ from one another in terms of how the signal is A. General OFDM signal model
structured. In particular, we shall see that the pilots used for In COFDM systems, the information is carried by a certain
channel estimation are differently positioned in LTE signals number Nc of orthogonal sinusoids, equally spaced with the k-
than in DVB ones, and that LTE signals also suffer (from a th carrier frequency expressed as fk = fc + k/Ts , where fc is
radarist perspective) from heavy power fluctuations. Numerous the center frequency, and all simultaneously transmitted over
papers thus explore the impact of LTE signal structure on radar the symbol duration Ts . These frequencies will be denoted
processing by studying its ambiguity function (e.g. [5]). In as subcarriers in the following. The basis generated by these
particular, ambiguities linked to synchronization signals and subcarriers for the m-th symbol can then be expressed as:
the aforementioned pilots have been studied in depth.
However, although many successfull drone detection exper- 
gk (t) = exp (2jπfk t) , 0 ≤ t < Ts
iments exist, in particular for DVB-T, only few papers present ψm,k (t) = gk (t), (1)
gk (t) = 0 elsewhere.
results on real LTE data. Work in [2] can be mentioned, where
the reference antenna pointing towards the transmitter and Let cm,k be the complex element of a given communication
surveillance antenna were placed on a top of a building, with alphabet (e.g. 16-QAM), transmitted in the m-th symbol on the
k-th subcarrier. The transmitted signal can then be expressed
as:
+∞
X c −1
NX
s(t) = cm,k ψm,k (t), (2)
m=−∞ k=0

which is usually implemented through a Fast Fourier Trans-

form. Finally, the transmitted signal includes a guard interval
Tu , also called Cyclic Prefix (CP), that is used to tackle
multipath reflections, which would otherwise degrade the
orthogonality of the signals. Each symbol thus consists of a
CP and of a useful duration corresponding to Ts . On this CP,
a portion of the end of the data symbol is copied. This induces
a periodicity of the signal, which will be used later. The basis
of the signal with guard interval is then expressed as:

gk (t) = exp (2jπfk t) , −Tu ≤ t < Ts
ψm,k (t) = gk (t),
gk (t) = 0 elsewhere.
(3) Fig. 1. LTE signal structure (issued from http://sharetechnote.com/). The
synchronization signals (PSS and SSS) are in blue. The CRS are in orange.
B. LTE signals structure PDCCH is in purple, PSCH in pink.
Although they share the same modulation, DVB and LTE
signals present some specificities. In this paper, we will
focus on Frequency Division Duplexing (between uplink and
downlink), even though Time Division Duplexing also exists.
LTE signals are decomposed in frames of duration 10ms, each
one being divided into 10 subframes. Each subframe is divided
into 2 slots of 0.5ms each. Figure 1 illustrates the general LTE
structure that is described in the following. A slot contains 6
or 7 symbols. Without loss of generality, we will consider the
case of a 7-symbol slot. Besides this temporal organization,
LTE signals are also divided in the time-frequency domain into
resource blocks. It is generally based on a set of 12 consecutive
subcarriers over 7 COFDM symbols.
LTE signals contain mainly two synchronization signals,
called Primary Synchronization Signal (PSS) and Secondary
Fig. 2. LTE CRS positions when two ports are simultaneously transmitting.
Synchronization Signal (SSS). They are helpful for both The CRS are in orange.
estimating the cell identities and the timing within a frame.
These signals are contained in the 6th (for the SSS) and 7th
(for the SSS) symbols of the 0th and 10th slots within a frame. polarization, plus 2 additional channels in dual polarization
They are carried by 62 subcarriers around the DC component. used to estimate the direct path (reference) signal. The radar
Another specificity of LTE signals is that the CP of the 1st processing used is quite similar to that used for DVB sig-
symbol of each slot can be slightly longer than for the other nals in [1], and consists of 4 main steps. First, the signals
symbols. For signals sampled at 30.72MHz, the CP can last are synchronized in both time and frequency, by using the
either 160 samples or 144 samples. Finally, the last specificity periodicity induced by the use of CP. The position within a
that we will account for is the pilots positioning. These are frame is given by correlation with PSS and SSS. In a second
usually refered to as Cell-specific Reference Signals (CRS) in time, the signals undergo a clutter rejection. This is done in
LTE framework. Their positions within a resource block de- the frequency domain, by estimating for each frequency the
pends on the number of transmitters. For one transmitter, they covariance matrix over the receiving channels. For the k-th
are localized on symbol 0 and symbol 4, every 7 subcarrier. subcarrier, the N × N covariance matrix R can be expressed
When a signal is transmitted on two polarizations, the CRS as:
of the two polarizations are present in the same symbol, but 1 X
Rk = y(k)y(k)H (4)
with a shift of 3 subcarriers. This is illustrated in Figure 2. TCP I
III. D ESCRIPTION OF THE RADAR PROCESSING where y ∈ CN is the vector of the data measured on the
To perform the detection of moving targets, we consider different channels for a symbol, and TCP I is the Coherent
a system composed of N = 12 receiving channels in V Processing Interval (CPI) over which the data are averaged.
 The inverse of this matrix is then applied on each channel
to obtain the clutter-canceled data:

ỹ(k) = R−1 y(k) (5)

The third step is the range, Doppler and angular compres-

sion of the clutter-canceled data. In our case, the reference
signal is directly the signal measured in the direction of
the direct path. This is because the decoding process in our
environment was complex, as we will see in section V. Finally,
the fourth and last step is the extraction on the processed
data. It is performed using a CFAR detector, estimating the
environment on a limited number of range-Doppler cells
surrounding the cell under test.
Fig. 4. Amplitude of the spectrogram on symbol 0 (dB). The FFT was
IV. E XPERIMENTAL SETUP calculated on 2048 sampled, in accordance with the 3GPP norm. The vertical
axis displays the frequency bins.
As already mentioned, we consider a system composed of
N = 12 receiving channels in V polarization to measure the
target, and 2 additional channels in dual polarization. The
data center frequency was fc = 1.87GHz, and the signals
were sampled at 25MHz. To preserve the orhtogonality of the
subcarriers, the signals were then resampled at 30.72MHz, in
accordance with the 3GPP norm.
The UAV used was a DJI Phantom 3, evolving in a rather
quiet area, without obstacles, at a velocity of about 10m/s.
The coherent integration time used was set to 125ms. Three
distinct transmitters were in the area. The relative positions of
the transmitters, the receivers, and the UAV are represented
in Figure 3. The UAV evolved in a more or less rectiligne
trajectory, going away from the receivers and back at it.
Figures 4 and 5 illustrate the data trafic on two symbols:
symbol 0 and symbol 1. Symbol 0 contains some CRS, Fig. 5. Amplitude of the spectrogram on symbol 1 (dB). The FFT was
whereas symbol 1 does not. It can be seen that there is a calculated on 2048 sampled, in accordance with the 3GPP norm. The vertical
huge difference in energy between the two symbols, symbol axis displays the frequency bins.
0 showing significantly more occupied resource elements than
symbol 1.
V. E IGENVALUES ANALYSIS
We begin the study of our results with an analysis of
the eigenvalues of the covariance matrix of the data in the
frequency domain R. Figure 6 shows the level of the eigen-
values estimated on symbol 0. In Figure 6, all the subcarriers
are displayed. A coarse observation shows that there are 3
preponderant eigenvalues (of level around 125dB, 110dB and
103dB), plus lower ones that are a little above the noise level.
This seems to indicate the presence of at least two transmitted
signals, and potentially other interferences. This made the
decoding of the signal of the direct path complicated, which
is why we used a measure of this direct path as the reference
signal in the range compression.
It can also be seen that the eigenvalues are strongly fluc-
tuating. In fact, this reflects that the spectral content is rep-
resentative e of three different components: CRS purely in H
Fig. 3. Scene configuration in a Cartesian grid. The three transmitters polarization, CRS purely in V polarization, and the transmitted
are represented by diamonds and denoted Tx, and the receiving array is
represented by a square and denoted Rx. The area in which the UAV flew is data, which are a mixture of the two polarizations. Indeed,
represented in orange. by zooming on a few subcarriers, a pattern can be observed.
Indeed, a peak seems to appear every 6 or 3 subcarriers,
which can be linked to the CRS structure, as they follow
the same frequency spacing (in the case of two transmitting
polarizations).
To confirm this statement, Figure 7 shows the same eigen-
values, but displayed only for the CRS corresponding to one
polarization (so every 6 subcarriers -see II-B). It can be
observed that the fluctuations are now much smoother.
By observing Figure 8, which represents the eigenvalues
estimated on symbol 1 (that does not contain CRS), a dif-
ferent behavior can be noted. This time, some high energy
burst can be observed (between 12 and 15MHz) on the two
highest eigenvalues. This can be explained by the strong power
variations that were observed in Figure 5, and the more regular
one observed in 4 (the presence of CRS guarantees a more
continuous power transmission).
From these observations, it can be assumed that depending
on the symbol(s) used, the range Doppler maps obtained
after clutter cancellation will behave differently. This will be
Fig. 7. Eigenvalues level (dB) estimated on symbol 0 in function of
examined in the next section. frequencies. Each color corresponds to a different eigenvalue. Only the
subcarriers corresponding to the CRS in polarization H are displayed.

Fig. 6. Eigenvalues level (dB) estimated on symbol 0 in function of frequen-

cies. Each color corresponds to a different eigenvalue. All the subcarriers are
displayed. A zoom on a few subcarriers is shown in the box surrounded by Fig. 8. Eigenvalues level (dB) estimated on symbol 1 in function of
the dotted lines. frequencies. All the subcarriers are displayed.

VI. D ETECTION RESULTS

The objective of our experiment was the detection of a UAV. corresponding respectively to Figures 9, 10 and 11. The
We will now proceed to analyse the outputs of the range detection obtained after applying the Constant False Alarm
Doppler compression on clutter-cancelled data. To confirm Rate detector with a threshold of 13dB are represented on
our earlier guess, we will perform this compression using the these figures. It should be noted that for a better visibility of
different symbols: on the CIT, we will first only keep the same the ambiguity function output (after clutter cancellation), the
symbols (e.g. all the symbols corresponding to the first one of surveillance and reference channel are shifted by 10 samples
the slot), and then use all the symbols available. We will then (so the direct path appears localized at the 10th sample).
analyse how it impacts the range Doppler maps and detection Besides from these maps, the detection results are summarized
results. for every different symbols in Table I, to enable a more global
As representative results, we will show the maps obtained disccussion. We recall that the symbols containg CRS are
using symbol 0 only, symbol 2 only, and all the symbols, symbol 0 and symbol 4. The target is detected when using
 TABLE I
D ETECTION RESULTS .

Symbol(s) Number of Target Target Target Target

used for of range Doppler θ SNR
compression plots (index) (index) (dB)
0 19 22 121 75 17.5
1 13 22 121 78 16
2 24 Missed Missed Missed 15.6
3 26 Missed Missed Missed 15.6
4 14 22 121 78 17.9
5 33 22 121 70 15.5
6 18 22 121 75 13.5
All 62 22 121 76 24.2

symbol 0, 1, 4, 5, 6 only, or when using all the symbols. It

is missed when using symbols 2 or 3 only. Even if the target
is detected when using some of the CRS-free symbols (5 and
6), its SNR is lower (ranging from 13.5 to 16dB), indicating
a degradation of the detection capacities. The highest SNR
is obtained by using all the symbols (24.2dB), followed by Fig. 9. Range Doppler map gain (dB) obtained using symbol 0. The detections
obtained after thresholding are represented by the red squares. The target
configuration using symbols containing the CRS (17.5 and position is indicated by the black cross.
17.9dB). However, the configuration using all the symbols
also shows the highest number of plots, while only one plot
corresponds to the target. This number is about 3 times higher
than the one obtained when using symbols containing the CRS.
This can be well visualized by comparing Figures 9 and 11
When analyzing the Range Doppler map in Figure 10,
obtained by using symbol 2, we can see the contrast is
decreased when compared to the one obtained using symbol
0 (Figure 9). This explains why the target is difficult to detect
in the CRS-free symbols.
The range of the detected points evolution with time is
presented in 12 and 13, where we used respectively all the
symbols or only symbol 0. Two different threshold values
are used in 12: 13dB or 18dB. The number of plots when
thresholding at 13dB is significantly higher than when using a
18dB threshold, which makes the target hard to be discernated.
At 18dB, the UAV trajectory can be guessed. When only
symbol 0 is used, the same trajectory can be observed when
using a 13dB threshold.
Summarizing these results, even if compressing the data us- Fig. 10. Range Doppler map gain (dB) obtained using symbol 2. The
ing all the symbols increases the SNR, it also leads to a hardly detections obtained after thresholding are represented by the red squares. The
target position is indicated by the black cross.
manageable number of plots. The use of symbols restricted
to those containing CRS leads to a slightly degraded SNR,
but also makes the extraction of the target significantly easier. reference signal. The analysis of the presented results seems
Finally, Figure 14 shows the final extraction of the target, using to indicate that, with the processing considered, compressing
the CRS containing symbols, after angular filtering. We can the data using only symbols containing the CRS leads to
see that only a few points of the trajectory are missing in the significantly improved detection capacities. This is mostly due
extracted plots, but they correspond to low Doppler instants, to the strong power fluctuations of the signal, making most of
for which the target is rejected along with the direct path. the symbols ”empty”, while the CRS containing ones always
contain at least the pilots. Using these symbols, the UAV
VII. C ONCLUSION
extraction was possible. Further work will include the pre-
We presented results of UAV detection using 4G-LTE-based sentation and discussion of other experimental configurations
passive radar. The environment in which the UAV evolved was (in particular other trajectories), and an analysis of limiting
challenging in terms of electro-magnetic environment, as more phenomena encountered when processing the data. A possible
than one signal was transmitted. The range compression was improvement would be the correct decoding of the direct path
still possible by using a measure of the direct path as the signal, making the reference used for compression cleaner and
hence improving the range compression.
R EFERENCES
[1] D. Poullin, “Passive Radar using COFDM (DAB or DVB-
T) Broadcasters as Opportunistic Illuminators,” in Digital Video.
London, United Kingdom: IntechOpen, 2010 [Online]. Available:
https://www.intechopen.com/chapters/8520 doi: 10.5772/8034.
[2] Dan, Yangpeng and Yi, Jianxin and Wan, Xianrong and Rao, Yunhua
and Wang, Benjing, “LTE-based passive radar for drone detection and
its experimental results,” The Journal of Engineering, October 2019, doi:
10.1049/joe.2019.0583.
[3] M. P. Jarabo-Amores et al., ”Drone detection feasibility with passive
radars,” 2018 15th European Radar Conference (EuRAD), Madrid,
Spain, 2018, pp. 313-316, doi: 10.23919/EuRAD.2018.8546549.
[4] M. Ummenhofer, L. C. Lavau, D. Cristallini and D. O’Hagan, ”UAV
Micro-Doppler Signature Analysis Using DVB-S Based Passive Radar,” Fig. 12. Range evolution with time obtained using all the symbols. The
2020 IEEE International Radar Conference (RADAR), Washington, DC, threshold used for detection was either 13dB or 18dB. The target range
USA, 2020, pp. 1007-1012, doi: 10.1109/RADAR42522.2020.9114747. evolution is visible when using 18dB. The plots corresponding to it are
[5] A. Evers, J. A. Jackson, “ Analysis of an LTE waveform for radar highlighted in purple.
applications,”. Proc. IEEE Radar Conf., Cincinnati, OH, USA, 2014,
DOI:10.1109/RADAR.2014.6875584.
[6] Raja Abdullah, R.S.A., Salah, A.A., Ismail, A., Hashim, F.H., Abdul
Rashid, N.E. and Abdul Aziz, N.H. (2016), “Experimental investigation
on target detection and tracking in passive radar using long-term evolu-
tion signal‘,” IET Radar Sonar Navig., vol. 10, pp 577-585, 2016,doi:
10.1049/iet-rsn.2015.0346.
[7] R. Maksymiuk, K. Abratkiewicz, P. Samczyński, M.Płotka, “Rényi
Entropy-Based Adaptive Integration Method for 5G-Based Passive
Radar Drone Detection,” Remote Sensing, vol. 14, no. 23, doi:
10.3390/rs14236146.

Fig. 13. Range evolution with time obtained using symbols 0. The threshold
used for detection was either 13dB. The target range evolution is visible when
using 18dB. The plots corresponding to it are highlighted in purple.

Fig. 11. Range Doppler map gain (dB) obtained using all the symbols. The
detections obtained after thresholding are represented by the red squares. The
target position is indicated by the black cross.

Fig. 14. Detected plots and ground truth of the UAV trajectory. The receiver
R is located at (0,0) (red square). The detected plots are indicated by the red
dots, and the ground truth is the black curve.