Wireless Networks (2022) 28:3537–3548

https://doi.org/10.1007/s11276-022-03010-z (0123456789().,-volV)(0123456789().,-volV)

ORIGINAL PAPER

Angle of arrival estimation in a multi-antenna software defined radio

system: impact of hardware and radio environment
Marcin Wachowiak1 • Pawel Kryszkiewicz1

Accepted: 16 May 2022 / Published online: 27 July 2022

 The Author(s) 2022

Abstract
Contemporary wireless communication transceivers can utilize their multiple antennas to improve positioning abilities.
Angle of Arrival (AoA) estimation utilizes shifts in phase between the signal of interest arriving at different receiving
antennas. Software Defined Radio (SDR) USRP B210 and GNU Radio implementing the Root Multiple Signal Classifi-
cation (Root-MUSIC) algorithm are used in this paper. However, this setup requires the consideration of errors caused by
hardware and radio environment. The impact of these effects is shown using an analytical model, assessed, mostly by
measurements, and relevant solutions are proposed. The hardware distortions are caused mostly by synchronization errors.
The intricacies of the most problematic phase synchronization are investigated with the wired setup showing, e.g., changes
with frequency or gain. Both synchronizations emitting calibration tone from the same radio front end, resulting in cross-
talk, or a separate one are tested. Recommendations enhancing the performance of the system and alleviating hardware
imperfections are provided in this paper. The accuracy of AoA estimation is degraded by multipath propagation or radio
interference. A signal processing scheme including filtering has been proposed. A set of over-the-air experiments assessed
the performance of the system. The presented and publicly available software package, scalable to a higher number of RX
antennas, enables straightforward implementation by other researchers.

Keywords Angle-of-Arrival  Direction-of-arrival  Phase synchronization  Software defined radio

1 Introduction for several services ranging from application-layer ser-

vices, e.g., driving assistance or indoor positioning to
The increasing demand for broadband wireless communi- improvement of network operation, e.g., using so-called
cations has resulted in a number of technologies being Radio Environment Maps [2, 3].
implemented in a contemporary wireless transceiver. One One of the basic schemes for location estimation using
of these is Massive Multiple-Input Multiple-Output multiple antennas is the Angle of Arrival (AoA) technique.
(mMIMO) [1] that allows for a significant network capacity Given a number of antenna elements spaced at regular
increase by utilization of multiple antennas at the Base intervals the angle of arrival can be discerned by observing
Stations (BSs). These antennas can also be used for other differences in the phase of received signals. Angle esti-
purposes, e.g., to allow for signal source positioning. The mation methods can be categorized into three groups:
availability of precise positioning information facilitates beamforming, subspace, and parametric [4]. Subspace
the prevalence of interactive forms of services or market- algorithms offer a compromise solution between accuracy
ing. One may imagine possibilities like assisted sightseeing and computational complexity. Among subspace methods,
in historical areas or museums, shopping racks with prox- the Root Multiple Signal Classification (Root-MUSIC)
imity-based advertising or assistance in commuting. algorithm provides the most accurate results, as presented
Overall, the angle-of-arrival-based positioning can be used in [4]. AoA estimation accuracy greatly depends on the
system’s hardware and radio conditions. The main hard-
& Pawel Kryszkiewicz ware problem encountered in AoA systems is the accuracy
[email protected] and stability of the radio front end phase synchronization.
1
Alignment in the frequency and time domains can be easily
Poznan University of Technology, Polanka 3, Poznan, Poland

123
3538 Wireless Networks (2022) 28:3537–3548

achieved by distributing a reference clock serving as a compensate phase differences by multiplying the stream of
Phased Locked Loop (PLL) input signal and gating pulses signal samples by a complex value representing the
marking specified moments in time. Due to the high scale required shift in the processing software. This approach,
of integration and multiple stages of signal processing or although simple, does not take into consideration the pos-
conversion, the resultant phases of end signals significantly sible drift of the coefficient. Moreover, there is no dis-
differ among front ends in the system. Moreover, manu- cussion on the validity of the measured coefficient, e.g., for
facturing inaccuracies and temperature dependency espe- a different device, carrier frequency, or gain.
cially affect the phase of the signal and may further It is also possible to synchronize the receiving front ends
contribute to the offset. In the case of wideband or agile without the need to reconfigure connections at RF ports.
Radio Frequency (RF) transceivers, the issue of phase When the SDR device is equipped with RF front ends with
synchronization becomes even more complicated as the two coherent RX channels, one may be utilized as input for
phase offset between RF front ends depends on the center the reference signal distributed with matched length cables.
frequency [5]. On the other hand, the radio environment Such an approach was presented in [11] with the use of
can constitute a significant source of AoA error. Multi-path National Instruments USRP 2954R. In this case, the RX
propagation causes a signal from a single source to arrive channel used for phase alignment is excluded from oper-
from many different directions. Moreover, it is possible ation in the AoA estimation phase, reducing its potential
that many signal sources will be active in the vicinity of the performance. Moreover, it is required that both RX chan-
AoA receiver causing interference and resulting in AoA nels of a given front end introduce a similar phase shift.
estimation error. The precise estimation of the phase shift difference
The introduction of Software-Defined Radio (SDR) requires precise cables and splitters to be used. Dissimi-
platforms has enabled rapid system prototyping and algo- larities between signal paths can contribute to an uncom-
rithm evaluation due to the flexibility of the signal pro- pensated accuracy error in this method. As before, an
cessing stage and extensive control over the radio front additional independent transmitter is required for syn-
ends. By using SDRs, the data stream of each receiving chronization. However, the calibration routine does not
(RX) or transmitting (TX) channel can be processed in a require any manual intervention.
personal computer serving as a central hub of the system. The previously mentioned white paper concerning a
There are several low-cost devices that enable quick AoA gr-doa package was more widely discussed in [12],
prototyping. In [6], a Texas Instruments development board where different hardware implementations were described.
consisting of 6 switched antennas is used in AoA estima- The setup consisting of multiple USRPs N210, each with 1
tion. Moreover, RTL-SDR modules with an additional RX and 1 TX channel, was presented with an unusual
clock distribution network can be used in this context [7]. calibration method. The phase synchronizing tone was
Also, there exist some off-the-shelf solutions that consist of inserted into the TX port of the device with matched cable
multiple RTL-SDR chipsets with an embedded clock dis- lengths. Due to the internal switching network between the
tribution system [8]. Even selected Wireless Local Area TX antenna port and RX channel, used normally for half-
Network (WLAN) access points with software modifica- duplex operation, a cross-talk occurs between both antenna
tions can be utilized to perform coarse AoA estimation [9]. ports. The tone inserted into the TX antenna port was
Still, probably the most popular SDR devices used in received at the RX side without the need to connect any-
prototyping and evaluation of wireless systems are the thing to the RX port or rearrange the wiring. The authors
Universal Software Radio Peripherals (USRPs) produced claim that the similarity of RX and TX signal paths result
by Ettus Research. in a repeatable phase shift across the device, providing
The popularity of SDR devices has brought about mul- sufficient calibration accuracy. It should be noted that the
tiple publications concerning the implementation of angle- USRP N210 clock used in the mentioned setup produces
of-arrival systems in various SDRs considering numerous phase ambiguity, which results in a random initial phase of
approaches to front end synchronization. Ettus Research the main Local Oscillator (LO) in a single device after each
has published a white paper which describes device syn- power-up or change of parameters. This means that the
chronization and showcases AoA estimation with the use of calibration result is not persistent and has to be performed
an X310 device [10]. The presented one-time calibration during runtime after the parameters have been set. What is
routine consists of attaching a signal source via a splitter to more, the accuracy of the calibration may be affected by
all RX channels simultaneously. Then, for a short period, manufacturing and component tolerance across signal
the phase difference between channels is measured, aver- paths and devices. Due to the cross-talk, an additional
aged, and saved to a file. Next, the wired setup is replaced transmitter is required to provide a synchronizing tone
by antennas to perform AoA measurements. The saved distributed across all devices.
coefficients representing the phase shift are applied to

123
Wireless Networks (2022) 28:3537–3548 3539

A simple, yet possibly less accurate method of front end software is easily scalable for a greater number of
synchronization is provided in [13]. The received signals antenna elements and is independent of the SDR device.
are visually aligned in an interactive software environment • An investigation of the impact of crosstalk (with the
by introducing an additional phase shift. It is done with the derivation of upper bounds), channel gains, time, and
reference signal connected to all RX ports via a splitter. In center frequency on the value of the calibration phase
addition, amplitude differences are noted. Whereas it might obtained for a wired synchronization of USRP B210
not be the most accurate method, it may be valid for quick front ends. Explanation based on the internal structure
alignment of phases and amplitudes of received signals. of the front end is given. The value of the calibration
Thermal effects associated with the heat up of SDR coefficient is compared for multiple USRP B210 units.
devices may introduce drifts or fluctuations of the phase • A proposal of solutions against noise, multipath, and RF
shift coefficient. In [14], the phase drift introduced by interference effects on AoA estimates in the constructed
hardware heat-up of a USRP X310 is measured over time. system. Applied averaging, band-pass filtering and
The calibration phase coefficient is shown to change over movement are verified and results are compared by
time. As such, delay from the last calibration routine can tests in a real environment.
increase phase error that affects the estimated AoA. The • An assessment of the overall accuracy of a setup with 2
work proposes an over the air calibration method using antennas using the Root-MUSIC AoA algorithm. A
periodic, pulsed signals. A single reference antenna is statistical characterization of the observed AoA errors
placed in front of the center of the array and transmits depending on the source of distortion.
pulses of the sine wave. Taking into account the calculated This paper is structured as follows: In Sect. 2 basics of
Euclidean distances from the reference antenna to RX AoA estimation are presented together with analytical
antennas, the expected phase shift introduced by propaga- models of synchronization errors. In Sect. 3 details of a
tion is compensated while calculating the calibration phase wired phase synchronization routine are discussed, pro-
coefficients among RX antennas. viding insight into different parameters affecting calibra-
In this work, an AoA implementation using a cheap tion. Finally, Sect. 4 concludes with a description of a field
SDR device, namely USRP B210, is presented. The main test setup and assessment of the accuracy of obtained AoA
focus is on factors limiting the accuracy of such a setup. An estimation results. The paper is concluded in Sect. 5.
analytical model of time, frequency, and phase error are
presented. While the time alignment is obtained via soft-
ware procedures embedded in the SDR’s driver and fre- 2 AoA estimation in a linear phased array
quency synchronization results from the utilization of a with synchronization imperfections
common LO, phase alignment has to be obtained sepa-
rately. Wired calibration using a single tone is used. Cali- A uniform linear array consists of N antennas regularly
bration using the same and separate TX is compared. separated by distance d as presented in Fig. 1 for N ¼ 5.
Measurements show, e.g., significant phase variation in The reference point of the geometry of coordinates
frequency and the utilized radio unit. The influence of [0, 0, 0] in Cartesian coordinates, is set to be the center of
thermal effects and the utilized RX gains are measured. All the array.
these results can be useful for other researchers planning to Let us consider a signal y(t) being transmitted with
implement AoA estimation using the B210 platform even if carrier frequency fc and complex envelope s(t), i.e.,
another synchronization method is used. Secondly, the
influence of the radio environment is considered. A proper yðtÞ ¼ sðtÞej2pfc t : ð1Þ
measurement procedure is suggested to combat multipath When distance d between array elements is much smaller
propagation effects. Moreover, a combination of filtering
and averaging is shown to increase robustness against radio
interference. The proposed routines for calibration and
angle-of-arrival estimation using the Root-MUSIC algo-
rithm are validated by measurements.
The main contribution of this paper is:
• Fully documented, open-source, and tested software for
AoA estimation in the GNU Radio environment. Phase
synchronization of an antenna array is obtained via a
wired setup within a short calibration period. The
Fig. 1 Uniform linear array with an incident plane wave

123
3540 Wireless Networks (2022) 28:3537–3548

than the distance from the signal source to the antenna 1X K

array, the incident electromagnetic wave can be assumed to Cx ¼ rðtk Þ rH ðtk Þ: ð6Þ
K k¼1
arrive at each antenna element from the same angle
denoted as h in the form of a plane wave. We assume that The calculated matrix is referenced as Cx – the sample
the wireless channel scales down the received signal correlation matrix. The obtained matrix serves as input to
amplitude, identically for each receiving antenna, by a the Root-MUSIC algorithm [15].
factor h. Signal y(t) arrives at the center of the array with
delay s0 . However, as a result of the array’s geometry, the 2.1 Front ends synchronization problem
signal arriving at the n-th antenna (n 2 f0; . . .; N  1g)
traveled the distance different by Depending on the utilized architecture, the receiving front
  ends can exhibit various synchronization errors between
N1
Dxn ¼  n d cos h: ð2Þ each other. This section is to exemplify how time, fre-
2
quency, and phase errors can influence the received signals
This causes the received signal at the n-th antenna to equal used for AoA estimation. The synchronization of front ends
  can be considered in three domains:
Dxn j2pfc ðts0 Dxcn Þ
rn ðtÞ ¼ hs t  s0  e ð3Þ Frequency: It can be assumed that the n  th receiving
c
radio front end uses in its mixer frequency fc þ Dfn , that is
where c denotes the speed of light. shifted by Dfn from fc . Most importantly, Dfn can vary
Assuming the reception delay difference for the n-th among front ends. This results in a modified form of (5) for
antenna Dxc n is much smaller than the symbol duration of the a received signal with frequency offset denoted as r~Fn ðt0 Þ:
complex envelope signal s(t), approximation
r~Fn ðt0 Þ ¼ s~ðt0 Þej2pDfn t ej2pð Þ
N1 d
  2 n k cos h : ð7Þ
Dxn
s t  s0  c  sðt  s0 Þ can be used. Therefore,
assuming the same carrier signal ej2pfc tjDu of phase shift It is visible that from the AoA estimation perspective, the
Du is used for down conversion at each front end and received signal phase will change by a factor of 2pDfn t,
sampling moment t0 is identical at each front end, the i.e., proportionally to the time and frequency difference.
received complex baseband signal at the n-th front end This is solved typically by sharing a reference clock signal
equals among the front ends.
Phase: Due to manufacturing inaccuracies and tolerance
r~n ðt0 Þ ¼ hsðt0  s0 Þej2pfc ðs0  c ÞjDu :
Dxn
ð4Þ
of components, the phase of the signal received by the
The common part of the signal arriving to each antenna can n  th radio differs among front ends. The phase shift can
be denoted as be visible both at the received signal path or the mixing
signal. Without loss of generality we can express it as a
s~ðtÞ ¼ hsðt  s0 Þej2pfc s0 jDu . By introducing it to (4)
and using (2) it is obtained: factor ejDun , specific for a given receiving front end,
modifying the locally generated signal used in the mixer.
r~n ðt0 Þ ¼ s~ðt0 Þej2pð Þ
N1 d
2 n k cos h : ð5Þ This results in a modified form of (5) for a received signal
with phase offset denoted as r~Pn ðt0 Þ:
This formula shows that in the case of perfect time and
r~Pn ðt0 Þ ¼ s~ðt0 ÞejDun ej2pð Þ
N1 d
frequency synchronization of all front ends, the phase of 2 n k cos h : ð8Þ
received signals differs only in phase depending on the
antenna index n and AoA h. The above formula can be The received signal varies among antennas by a phase shift
extended for a higher number of incoming signals (with Dun . While its value can range from 0 to 2p, it is typically
different AoAs) and the presence of white noise. fixed or slowly varying in time. Therefore, a common
Let us assume that the received signal is simultaneously approach is to calibrate the array estimating N  1 coeffi-
sampled across all N antenna elements at time instance tk . cients. One of the phase shifts, e.g., Dun~ can be embedded
The obtained data is denoted as a column vector of N as a common factor into s~ðt0 Þ, requiring to estimate phase
samples rðtk Þ :¼ ½r0 ðtk Þ r1 ðtk Þ . . . rN1 ðtk ÞT . Multiple shift difference ejðDun Dun~Þ for
Signal Classification (MUSIC) algorithms exploit proper- n 2 f0; . . .; n~  1; n~ þ 1; . . .; N  1g. For the n-th~ antenna
ties of array covariance matrix Rx , which can be estimated the coefficient is assumed to equal 1. Phase synchroniza-
using different approaches influencing the accuracy of AoA tion of the front ends is achieved typically with post pro-
algorithms [15]. One method is to estimate the matrix from cessing where the signal received from each antenna is
an average of K data snapshots as: multiplied by the conjugation of the estimated coefficient.

123
Wireless Networks (2022) 28:3537–3548 3541

Time: The received signals at all front ends should be identical configurations. The constructed radio system is
sampled at the same time instance. Moreover, this align- controlled via GNU Radio software where necessary
ment should not be lost during signal processing and pro- functions and routines are implemented. The signal pro-
visioning to the high-level software in SDR. Assuming cessing flowchart responsible for phase measurements can
there is a sampling moment shift at the n-th front end Dtn be seen in Fig. 2. The calibration was performed with a
the received signal equals single tone, of a frequency 50 kHz higher than the carrier
frequency, fed to the RX channels via a splitter and mat-
r~Tn ðt0 Þ ¼ s~ðt0 þ Dtn Þej2pfc Dtn ej2pð Þ
N1 d
2 n k cos h : ð9Þ
ched length cables. The test tone shift was performed in
If the time difference Dtn is much smaller than the trans- order to prevent LO leakage providing additional distortion
mitted signal symbol interval, i.e., s~ðt0 þ Dtn Þ  s~ðt0 Þ, and of the observed tone. An additional attenuator of 30 dB
not changing in time, the synchronization error causes only mounted before the splitter is a safety measure to ensure
a constant phase shift 2pfc Dtn that can be estimated and that the input signal does not exceed the maximum toler-
removed as part of phase synchronization. If this condition able power level of USRP B210 and prevents any damage
is not met, the complex signal envelope s~ðt0 þ Dtn Þ will to the RX chain. In the software domain, the signal is then
differ among front ends, preventing successful AoA esti- band-pass filtered by a Finite Impulse Response (FIR) fil-
mation. Therefore, coarse sampling time alignment is a ter, with a pass band of 45-55 kHz and 10 kHz transition
must. It is typically accomplished by a 1 Pulse Per Second width. Next, the phase difference is calculated and aver-
(PPS) signal that is distributed among devices and serves as aged by a Moving Average Filter (MAV). In the end, the
a gating pulse that resets timers. This signal is also avail- obtained stream is decimated and saved to a file. If not
able in most Global Positioning System (GPS) receivers. stated differently, the center frequency is set to 2.44 GHz.
As mentioned above, fine time synchronization is obtained The phase synchronization routine was examined for two
as part of the phase synchronization process. calibration source configurations. One with a signal from
Taking all synchronization errors (7–9) into considera- the TX of the receiving USRP (as shown in Fig. 3a) and the
tion, the final model of the received signals can be second with a signal provided by another USRP B210 (as
expressed as: shown in Fig. 3b). The sampling rate of both TX and RX is
set to 1 MSPS.
r~n ðt0 Þ ¼ s~ðt0 þ Dtn Þej2pðDfn tfc Dtn ÞjDun ej2pð Þ
N1 d
2 n k cos h :
ð10Þ 3.1 Comparison of wired synchronization setups
Notice that all the errors are visible as changes in the It was initially assumed that a single USRP B210 will be
received signal phase. As AoA estimation is based on sufficient to create a self-contained system with two
differences of received signals phases, this can cause a receiving elements and a calibration source utilizing one of
significant AoA error. Therefore, an important process in the TX channels. However, this setup is prone to errors
AoA estimation is the provisioning of the coarse syn- introduced by crosstalk between adjacent RX and TX
chronization among RX front ends and fine synchroniza- channels. The crosstalk in USRP B210 is a result of its
tion by proper phase rotation. While a fixed phase architecture as visible in Fig. 3a. The switches present in
correcting coefficient is desired, it is possible that its the signal chain have limited isolation and introduce cou-
measurement has to be repeated periodically or after pling between TX and the closest RX channel of SDR. The
receiver reconfiguration. This will be discussed for USRP calibration signal received by the front end with crosstalk
B210 based on measurements in Sec. 3.
can be expressed as: r c ðtÞ ¼ Aej2p ft þ Bej2p ftju , where
Aej2p ft is the signal received through the RX port (wanted).
The second component is the leaked signal (unwanted) that
3 Calibration Phase Measurements for USRP
B210

Phase calibration and measurements were performed with

the USRP B210 SDR platform. In this device time and
frequency synchronization of channels within a single unit
is provided by a shared clock. In the case of multiple
devices, the synchronization may be provided by reference
clock and a pulse-per-second signal being fed to each SDR.
The architecture of B210 guarantees that the phase shift
between channels remains similar between runs for Fig. 2 Phase shift measurement flowchart

123
3542 Wireless Networks (2022) 28:3537–3548

Fig. 3 Diagrams of USRP B210

wired phase calibration setups

(a) (b)

is shifted in phase by u in regard to the former. Without

loss of generality, it is assumed that A and B are ampli-
tudes, i.e., non-negative real numbers. After the down-
conversion, the carrier frequency component is removed,
c
giving rbb ðtÞ ¼ A þ Beju . The difference in phase is max-
imal when Beju is perpendicular to A in the polar repre-
sentation of complex numbers. As both A and B are non-
negative real numbers and A [ B, the maximal phase shift
introduced by B is when it is of purely complex value, i.e.,
jB or þjB. The maximal phase error can then be calcu-
lated based on the formed right-angled triangle as:
 
Dw ¼ arctan BA . When B\\A, the value of arctangent
can be approximated with the Maclaurin-Taylor series for Fig. 4 Accuracy of phase calibration in regard to attenuation attached
arctangent, resulting in: Dw  BA. As such, the worst case, to TX source
crosstalk-originating phase estimation error can be
approximated by a square root of the ratio of crosstalk and power depends on the utilized TX channel, both the
wanted signals powers. obtained bounds and the measured phase differ. However,
To verify the extent of crosstalk impact on the accuracy in both cases, the measured points are within the derived
of phase calibration, a series of measurements were per- bounds.
formed. For RX gains of 40 dB and TX gain of 45 dB the In wired setups it is common to use attenuators as a
crosstalk magnitude, i.e., B, was measured with all antenna safety measure to prevent any damage to the RX signal
ports of the same channel being terminated with a matched chain, the obtained results show it can lead to a significant
impedance load (see Fig. 3a). Next, the reference value of phase calibration error, e.g., a 30 dB attenuator introduces a
A was measured without the attenuator inserted between phase error of approximately 2-5 .
TX and RX2 ports. In this configuration, A is much greater Due to the observed levels of crosstalk and its influence
than B, allowing the crosstalk component to be omitted. on phase calibration, consecutive measurements are per-
Finally, the standard phase calibration routine was exe- formed with a separate USRP B210 operating as a refer-
cuted for a few values of inserted attenuation in the TX ence TX as depicted in Fig. 3b.
path. Given that for constant gains of both TX and RX of
USRP B210, the crosstalk level is also constant, increasing 3.2 Phase-time dependency
the value of attenuation reduced the desired signal, i.e., A,
allowing us to investigate the calibration accuracy for Here the phase correction coefficient was observed for a
various signal-to-crosstalk ratios. The results of the mea- long period of time. The test lasted around a few hours. RX
surement are presented in Fig. 4. The value of A for gains were 10 dB and TX gain was set to 50 dB. The
attenuation greater than 0 dB is calculated using the value additional safety attenuator value was 30 dB. Measure-
of the introduced attenuation. ments were performed for previously unpowered devices to
First, the measured phase difference obtained for observe the extent of coefficient drift. The collected data
transmission from a separate TX is invariant from the was smoothed by a MAV filter of the length of 100k
attenuation value (a series denoted as Sep. TX), the samples. The results presented in Fig. 5 show the initial
increasing attenuation causes an increasing phase error for drift caused by the heat up of the device and the distribu-
self-calibration configurations. Although the crosstalk tion of the measured values after stabilization. It can be

123
Wireless Networks (2022) 28:3537–3548 3543

Fig. 5 Change of phase shift correction coefficient value in wired setup over time

seen that the distribution of the phase difference is Gaus- response of the device is very stable over multiple sweeps.
sian once stabilized with a mean value of 8:9 and a While the phase difference among channels significantly
standard deviation of 0:05 . The initial drift during the first depends on the frequency and is fixed in time, one could
two hours of operation accounts for a change of the mean consider storing its values and using them when needed.
by approximately 0.07 , which is relatively small com- One might want to use the measured value for another
pared to the overall averaged value and may be considered B210 device.
negligible. Aiming to understand how the phase-frequency char-
acteristic varies between B210 devices, the frequency
3.3 Phase-frequency dependency sweep was performed for 6 available devices, resulting in
Fig. 7. From this plot it is visible that the phase-frequency
To analyze the dependence of the phase shift on carrier characteristic is individual for each device. Small differ-
frequency, additional measurements were performed ences introduced in integrated circuits and utilized elec-
spanning the range from 350 MHz to 6 GHz in 50 MHz tronic components result in significant, from the AoA
steps. It consisted of collecting the phase difference values estimation point of view, differences in the signal phase
over 10 s with MAV of 100k length for each carrier fre- rotation. There is no universal trend and it is hard to discern
quency. The procedure was repeated in 50 sweeps to check common features. The overall phase difference can be
if there is any variance between repeated measurements for contained mostly within the ð10 ; 20 Þ range. While the
the same frequency. Figure 6 depicts the average of all 50 results of our measurements can potentially be used instead
sweeps and 95% confidence intervals calculated over the of calibration measurements for USRP B210, it can result
means of 10 s measurements for a given frequency. It can in an AoA estimation error increased by around 10-20 in
be seen that the phase difference does not exhibit any some cases. Another interesting phenomenon are the steep
monotonic dependency on the center frequency. This changes in phase at the frequencies of 2.2 and 4 GHz.
irregularity is probably due to discrete components used in Those discontinuities originate from the switching of
the signal chain and their combined diverse phase charac-
teristics at a given frequency. The confidence interval
overlaps the mean, meaning that the phase-frequency

Fig. 6 Average of multiple frequency sweeps of phase shift coeffi-

cient value Fig. 7 Phase-frequency characteristic of multiple USRP B210 devices

123
3544 Wireless Networks (2022) 28:3537–3548

discrete component signal chains in USRP B210 due to appropriately shifted in phase. The corrected signals were
wideband frequency matching [16]. then correlated obtaining a correlation matrix according to
(6). In these measurements, the snapshot size K was set to
3.4 Phase-gain dependency 1024, which corresponds to the correlation matrix aver-
aging length. For the sample rate of 1 MSPS around 976
In order to investigate the influence of RX gain settings, the matrices per second are generated and a new AoA estimate
phase measurements were performed for a range of gain is obtained approximately every 1 ms. Next, each corre-
values. First, the gain of RX channel 0 has been fixed to 20 lation matrix is processed by the Root-MUSIC algorithm,
or 37 dB while the gain of the second channel was changed returning a single value of the estimated AoA, which is
over the whole range, i.e., from 0 to 76 dB. The resultant saved to a file.
phase difference significantly changes in the whole range
from 180 to 180 , as shown in Fig. 8a. The steep phase 4.1 Results
changes have been linked to switches of gain of Integrated
Circuits (ICs) of the RF chipset present in the USRP B210, The Angle of Arrival system performance was examined
e.g., Low Noise Amplifier (LNA) and mixer, as depicted in for various environmental conditions. Initially, the transmit
Fig. 3b. The distribution of the overall RX gain among IC node was placed indoors at an angle of 90 , 2 m from the
can be found in [17]. The plots representing the integrated array. Angle estimation was performed for 60 s, resulting
LNA (iLNA) and mixer gains were added to Fig. 8a to in about 58k of AoA values. The RX gains were set to 60
illustrate their correlation with the measured phase differ- dB and TX of the source to 50 dB. Figure 10 compares
ence. On the other hand, when both RX channels have AoA error distribution under various conditions obtained
identical gains, the change of phase difference is much after removing the mean value of AoA for each setup.
smaller, although still correlated with internal IC gain Figure 10a and b show the influence of RF interference
values, as shown in Fig. 8b. Throughout the whole sweep and how AoA estimation can be improved by using band-
the value changes at most by 1.8 . pass filtering, respectively. The radio interference was
generated by turning on a WLAN 802.11 access point
nearby operating at the carrier frequency of 2.442 GHz (7th
4 SDR-based angle of arrival estimation channel) and enforcing high traffic in the network. The
traffic was generated by streaming a Youtube video to the
The measured calibration phase shift allows for testing the smartphone to resemble a realistic scenario. The mobile
angle of arrival estimation performance. The utilized two- device was placed approximately at the same distance to
antenna setup and the signal processing flowchart are the array as the main source. The interference setup is
shown in Fig. 9. The transmitter was a separate SDR uti- visualized in Fig. 11.
lizing a single TX antenna generating a continuous wave. The utilized filter had the same parameters as in the
The constructed radio system was controlled via the GNU calibration tests in Sect. 3. AoA error distributions prove
Radio software environment. First, the signal was band- that band-pass filtering is absolutely necessary to limit the
pass filtered. Next, given the measured calibration coeffi- interference from overlapping bands in busy wireless
cient Du, the received signal over the second channel was channels. The AoA error introduced by other transmissions

(a) (b)
Fig. 8 Phase dependence on the gain of RX front ends

123
Wireless Networks (2022) 28:3537–3548 3545

Fig. 9 Angle of Arrival

measurement system

(a) (b)

(a) (b)

(c) (d)

Fig. 10 Probability distribution of AoA error in various conditions

source of distortion at a given time. The results, presented

in Fig. 10c and d, were obtained with applied band-pass
filtering and low RF activity in the utilized frequency band.
The movement was introduced by a person walking around
the system without crossing the Line-of-Sight between TX
and RX antennas. The AoA estimation error introduced by
movement dominates other sources and ranges between
10 with a standard deviation of 2.2 . This shows that
multi-path propagation has a significant impact on AoA
Fig. 11 RF interference setup estimation and a proper averaging of obtained AoA values
and measurement procedure design is needed to combat
spreads over the whole range of estimation. The standard
this error.
deviation of angle estimation without filtering equals
Therefore, measurements in an outdoor scenario were
approximately 8 . With the addition of filtering the stan-
carried with both transmitter and receiver distanced from
dard deviation is decreased to only 0.3 .
the buildings (by at least 10 m) and ground (by using 2m-
Subsequent measurement was aimed at investigating
high tripods). Moreover, the transmitter used a directional
how movement distorts the Angle of Arrival estimation.
antenna. The applied measures were to limit the total
The RF interference was disabled in order to show a single
power of non-Line-of-Sight components. RF interference

123
3546 Wireless Networks (2022) 28:3537–3548

Fig. 12 Accuracy of the AoA system in field tests

was not generated as the expected conclusions are the same maximum time the target is at a given AoA to utilize all
as in the indoor scenario. The signal source has been placed variations of multipath propagation.
at angles 90 and 135 at a distance of 5 m. The setup
operated at the frequency of 2.44 GHz with active band-
pass filtering. The first series of measurements were per- 5 Conclusions
formed for the source antenna in a fixed position. Then the
source was moved slightly forward and backward main- It is possible to build an accurate AoA estimator using a
taining the angle to modify potential fadings. cheap SDR platform. While USRP B210 provides accurate
The obtained results are shown in Fig. 12. Here, the enough time and frequency synchronization, phase syn-
AoA error was calculated with respect to the known source chronization has to be implemented by the user. Wired
location. First, outdoor measurements with movement calibration is suggested. While the obtained correction
show a smaller standard deviation than indoor measure- coefficient is relatively stable in time, it requires recali-
ments. This proves that if moving elements of the envi- bration after changing the front end setup, e.g., carrier
ronment are more distanced, as in our outdoor setup, the frequency, or when changing the utilized device. While
signal reflected from them has a smaller influence on the measuring AoA using the Root-MUSIC algorithm in a real
total received signal. Additionally, a comparison of static environment it is recommended to employ filtering to
and moving measurements in the outdoor scenario shows remove adjacent radio interference and use AoA averaging
that AoA estimation can generally be improved by changes and movement of RX or TX devices to cope with multipath
in multi-path propagation. While the static measurements propagation. This should result in a 2 antenna AoA esti-
exhibit a much lower variance of the result, they are biased mation error smaller than 10 in most environments,
by a significant error compared to the series with move- enabling, e.g., short-range coarse positioning or algorithms
ment, i.e., the mean error is on average higher. Moving the enhancing network performance. However, the obtained
antenna helps to visualize that the multi-path effect has a accuracy depends on many factors including radio propa-
crucial impact on the performance of the AoA estimation gation conditions. Moreover, it can be improved by
system. It can be noted that the AoA error of static mea- increasing the number of utilized antennas [18]. Finally,
surement constitutes a part of the distribution obtained for the positioning error depends on the AoA error in addition
the measurements with movements. Therefore, it can be to the utilized positioning algorithm or the number of AoA-
recommended to perform AoA estimation in the environ- reporting nodes [19]. The implemented AoA estimation
ment with movement, both indoors and outdoors, providing software is publicly available at [20].
a sufficiently long AoA estimates averaging period. The
Acknowledgements This research was funded by the Polish National
recommended averaging duration should span the
Science Centre, project no. 2021/41/B/ST7/00136. For the purpose of
Open Access, the author has applied a CC-BY public copyright

123
Wireless Networks (2022) 28:3537–3548 3547

licence to any Author Accepted Manuscript (AAM) version arising 2018 17th RoEduNet Conference: Networking in Education and
from this submission. Research (RoEduNet), pp. 1–6.
12. Pagadarai, S., & Collins, T. F. (2017). GR-DOA: Direction
Data availability The datasets generated during the current study are Finding in GNU-Radio, Available at https://www.gnuradio.org/
available from the corresponding author on reasonable request. grcon/grcon17/presentations/gr_doa_gnuradio_direction_finding/
.
13. Jokinen, M., Sonkki, M., & Salonen, E. (2017). Phased Antenna
Declarations Array Implementation with USRP. In IEEE Globecom Work-
shops (GC Wkshps), 2017, 1–5.
Conflict of interest The authors declare that they have no known 14. Monfared, S., Nguyen, T., Vorst, T. V. d., Doncker, P. D., &
competing financial interests or personal relationships that could have Horlin, F. (2020). Experimental Demonstration of AoA Estima-
appeared to influence the work reported in this paper. tion Uncertainty for IoT Sensor Networks. In 2020 IEEE 91st
Vehicular Technology Conference (VTC2020-Spring), pp. 1–5.
Open Access This article is licensed under a Creative Commons 15. Trees, H. L. V. (2002) Parameter Estimation II. Wiley, ch. 9,
Attribution 4.0 International License, which permits use, sharing, pp. 1139–1317.
adaptation, distribution and reproduction in any medium or format, as 16. Ettus Reseach A National Instruments Brand, Band selection
long as you give appropriate credit to the original author(s) and the switching frequencies of USRP B210. Available at https://github.
source, provide a link to the Creative Commons licence, and indicate com/EttusResearch/uhd/blob/master/host/lib/usrp/b200/b200_
if changes were made. The images or other third party material in this impl.cpp.
article are included in the article’s Creative Commons licence, unless 17. Analog Devices, ‘‘AD936x gain switching tables in .csv format,’’
indicated otherwise in a credit line to the material. If material is not Available at https://ez.analog.com/wide-band-rf-transceivers/
included in the article’s Creative Commons licence and your intended design-support/w/documents/10071/gain-tables-in-csv-format.
use is not permitted by statutory regulation or exceeds the permitted 18. Gentilho, T. A. Edno., Scalassara, P. R. (2020). Direction-of-
use, you will need to obtain permission directly from the copyright arrival estimation methods A performance-complexity tradeoff
holder. To view a copy of this licence, visit http://creativecommons. perspective. Springer Journal of Signal Processing Systems, 92,
org/licenses/by/4.0/. D239–256
19. Shao, H.-J., Zhang, X.-P., & Wang, Z. (2014). Efficient closed-
form algorithms for Aoa based self-localization of sensor nodes
References using auxiliary variables. IEEE Transactions on Signal Pro-
cessing, 62(10), 2580–2594.
1. Larsson, E. G., Edfors, O., Tufvesson, F., & Marzetta, T. L. 20. Wachowiak, M. Github repository containing gr-aoa module,
(2014). Massive MIMO for next generation wireless systems. with signal processing blocks and flowcharts used within this
IEEE Communications Magazine, 52(2), 186–195. work. Available at https://github.com/MarcinWachowiak/GNU-
2. Hoffmann, M., & Bogucka, H. (2019). Localization Techniques Radio-USRP-AoA.
for 5G Radio Environment Maps. In Cognitive Radio-Oriented
Wireless Networks (pp. 232–246). Cham: Springer. Publisher’s Note Springer Nature remains neutral with regard to
3. del Peral-Rosado, J. A., Raulefs, R., López-Salcedo, J. A., & jurisdictional claims in published maps and institutional affiliations.
Seco-Granados, G. (2018). Survey of cellular mobile radio
localization methods: From 1g to 5g. IEEE Communications
Surveys Tutorials, 20(2), 1124–1148. Marcin Wachowiak (IEEE Stu-
4. Theodoridis, S., & Chellappa, R. (2013). Academic Press Library dent Member) received the B.S
in Signal Processing, Volume 3: Array and Statistical Signal in Electronics and Telecommu-
Processing, 1st ed. USA: Academic Press. nications from the Poznan
5. Analog Devices, FMComms5 Phase Synchronization. Available University of Technology
at https://wiki.analog.com/resources/eval/user-guides/ad- (PUT), Poland, in 2021. He is
fmcomms5-ebz/phase-sync. currently pursuing a Master’s
6. Hajiakhondi-Meybodi, Z., Salimibeni, M., Plataniotis, K. N., & degree in the same field under
Mohammadi, A. (2020). Bluetooth low energy-based angle of the supervision dr Kryszkie-
arrival estimation via switch antenna array for indoor localiza- wicz. His research interests
tion. pp. 1–6 include software-defined radio
7. BniLam, N., Joosens, D., Steckel, & J., Weyn, M. (2019). Low and multiantenna systems.
cost Aoa unit for IoT applications, pp. 1–5
8. Dai, Z., He, Y., Tran, V., Trigoni, N., & Markham, A. (2022).
Deepaoanet: Learning angle of arrival from software defined
radios with deep neural networks. IEEE Access, 10, 3164–3176.
9. Schüssel, M.: Angle of arrival estimation using wifi and smart-
phones, pp. 1–4.
10. Ettus Reseach A National Instruments Brand, Phase Synchro-
nization Capability of TwinRX Daughterboards and DoA Esti-
mation. 2016, Available at https://github.com/EttusResearch/gr-
doa/blob/master/docs/whitepaper/doa_whitepaper.pdf.
11. Rares, B., Codau, C., Pastrav, A., Palade, T., Hedesiu, H.,
Balauta, B., & Puschita, E.(2018). ‘‘Experimental Evaluation of
AoA Algorithms using NI USRP Software Defined Radios. In

123
3548 Wireless Networks (2022) 28:3537–3548

Pawel Kryszkiewicz (IEEE His main fields of interest are problems concerning the physical layer
Senior Member) received the of multicarrier and multiantenna systems, green communications, and
M.Sc. and Ph.D. degrees Dynamic Spectrum Access. Dr Kryszkiewicz is the author of over 100
(Hons.) in Telecommunications research papers, 2 book chapters and 1 book.
from the Poznan University of
Technology (PUT), Poland, in
2010 and 2015, respectively. He
is currently an Assistant Pro-
fessor with the Institute of
Radiocommunciations, PUT.
Moreover, dr Kryszkiewicz is
the co-founder and the Techni-
cal Director of RIMEDO Labs,
the spin-off from PUT. He has
been involved in a number of
national and international (including EU-funded) research projects.

123