An Experimental mmWave Channel Model for UAV-to-UAV Communications
Michele Polese, Lorenzo Bertizzolo, Leonardo Bonati, Abhimanyu Gosain, Tommaso Melodia
AAn Experimental mmWave Channel Modelfor UAV-to-UAV Communications
Michele Polese, Lorenzo Bertizzolo, Leonardo BonatiAbhimanyu Gosain, Tommaso Melodia
Institute for the Wireless Internet of Things, Northeastern University, Boston, MA 02115, USAEmail: {m.polese, bertizzolo.l, bonati.l, a.gosain, t.melodia}@northeastern.edu
ABSTRACT
Unmanned Aerial Vehicle (UAV) networks can provide aresilient communication infrastructure to enhance terres-trial networks in case of traffic spikes or disaster scenar-ios. However, to be able to do so, they need to be basedon high-bandwidth wireless technologies for both radio ac-cess and backhaul. With this respect, the millimeter wave(mmWave) spectrum represents an enticing solution, since itprovides large chunks of untapped spectrum that can enableultra-high data-rates for aerial platforms. Aerial mmWavechannels, however, experience characteristics that are sig-nificantly different from terrestrial deployments in the samefrequency bands. As of today, mmWave aerial channels havenot been extensively studied and modeled. Specifically, thecombination of UAV micro-mobility (because of imprecisionsin the control loop, and external factors including wind) andthe highly directional mmWave transmissions require ad hocmodels to accurately capture the performance of UAV de-ployments. To fill this gap, we propose an empirical propaga-tion loss model for UAV-to-UAV communications at 60 GHz,based on an extensive aerial measurement campaign con-ducted with the Facebook Terragraph channel sounders. Wecompare it with 3GPP channel models and make the mea-surement dataset publicly available.
KEYWORDS mmWave, UAV, Cellular Networks, 60 GHz, Propagation.
ACM Reference Format:
Michele Polese, Lorenzo Bertizzolo, Leonardo Bonati and Abhi-manyu Gosain, Tommaso Melodia. 2020. An Experimental mmWave
Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copiesare not made or distributed for profit or commercial advantage and thatcopies bear this notice and the full citation on the first page. Copyrightsfor components of this work owned by others than the author(s) mustbe honored. Abstracting with credit is permitted. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee. Request permissions from [email protected]. mmNets’20, September 25, 2020, London, UK © 2020 Copyright held by the owner/author(s). Publication rights licensedto ACM.ACM ISBN xxxxxxxxxxxxxxxxxxxxxxxxx...$15.00https://doi.org/10.1145/xxxxxxx.xxxxxxxx
Channel Model for UAV-to-UAV Communications. In
ACM, New York, NY, USA,7 pages. https://doi.org/10.1145/xxxxxxx.xxxxxxxx
This paper has been accepted at 4th ACM Workshop on Millimeter-wave Networks and Sensing Systems (mmNets’20).Copyright may be transferred without notice.
Unmanned aerial systems are promising technological en-ablers for the wireless industry, as they provide an effectiveand inexpensive solution to temporarily connect groundusers in the absence of terrestrial infrastructure [4–6]. Inthis domain, a key research challenge is how to provide high-capacity robust backhaul and inter-Unmanned Aerial Vehicle(UAV) connectivity in flying platforms. Fiber optic backhaul,typical of terrestrial infrastructure, is not a feasible solutionfor UAVs, thus, both the radio access and the backhaul exploitwireless links.However, combining access and backhaul on the samewireless interface introduces tight data-rate and latency re-quirements to the underlying communication technology.This problem is exacerbated when multiple UAVs rely oneach other for data forwarding, in an aerial multi-hop fash-ion. While traditional sub-6 GHz technologies are unfit forhigh-load traffic aggregation, the millimeter wave (mmWave)spectrum can be a unified solution for fully-wireless nodes of-fering unprecedented bandwidth. Thus, at mmWaves, a UAVnetwork can use the same wireless technology to provideconnectivity to ground users, communicate with neighbor-ing UAVs, and relay data traffic toward the closest groundtower, in an “integrated access and backhaul” fashion [13].Even though recent 5G standard specifications alreadyenvision the use of mmWaves [1], and some previous workstudied their propagation for ground deployments [2, 15], theaerial wireless channel at mmWave frequencies has not beenextensively characterized yet. Indeed, mmWaves may affectthe communication quality in aerial scenarios differentlyfrom ground deployments, for the following reasons: (i) Somefrequencies in the mmWave spectrum (e.g., the 60 GHz band)suffer from oxygen absorption. As these atmospheric condi-tions change with the deployment height, high-altitude aerialscenarios might differ from ground deployments; (ii) even a r X i v : . [ c s . N I] A ug hen Global Positioning System (GPS)-locked, the inaccu-racy of UAV’s on-board sensors may lead to slight horizontal(yaw) and vertical (throttle) drone fluctuations. Given thatmmWaves exploit highly directional communications, thesefluctuations might severely deteriorate the channel qual-ity [5]; (iii) when UAVs fly in harsh wind conditions, theylean forward/backward (pitch) and sideways (roll) to coun-terbalance the wind force. Tilting a UAV-mounted mmWaveradio might compromise the link quality by changing thebest beam path or the radios’ polarization [12]; (iv) on-boardmmWave radios are often mounted either above or belowthe main UAV’s frame structure. The frame size and its mate-rial, together with the spinning propellers introduce ambientnoise right where the headset is mounted. On-board batteries,radio control, and circuitry non-linearity exacerbate this ef-fect known as “airframe shadowing” [17]. For these reasons,the performance evaluation of mmWave UAV-to-UAV com-munications cannot be based upon ground-tailored mmWavechannel studies and demands the development of dedicatedair-to-air (A2A) propagation and fading models. Prior workhas focused on analytical or ray-tracing approaches [7, 9–11], which have not been validated through experimentalmeasurement campaigns.In this paper, we propose an empirical propagation lossmodel for A2A communications at 60 GHz. It is based on anextensive measurement campaign, with more than 3 daysof flight experiments, and the Facebook Terragraph channelsounders [8] mounted on two UAVs. The measurements vali-date the empirical model in a wide range of flying heights(6 −
15 m) and distances (6 −
40 m), and show that, in theconsidered range, the path loss does not have an explicitdependence on the UAV height. Moreover, we compare thepath loss curve with 3GPP channel models, and, using thesame measurement campaign, we characterize the impact ofa sub-optimal beam selection on the link budget.
To the bestof our knowledge, this is the first empirical A2A propagationmodel for the 60 GHz band modeling the impact of the UAVsmicro-mobility on the channel.
Last, we publicly release thecollected measurements’ dataset and analysis scripts to thecommunity. The remainder of the paper is organized as follows. Sec. 2describes the measurement campaign. Sec. 3 analyzes empir-ical path loss fits on the collected data and compares themwith terrestrial channel models. Finally, Sec. 4 concludes thepaper. https://github.com/wineslab/uav-to-uav-60-ghz-channel-model For our measurements, we employed two DJI M600 as UAVsand two Facebook Terragraph mmWave radios configured aschannel sounders with beam scanning capabilities. Each DJIM600 mounts an on-board powered Intel NUC computer toperform flight control tasks, while the Facebook Terragraphradios are powered from the ground and coordinated in theirchannel sounding procedures by a ground host controller.The hardware and software schematics of the mmWave-enabled UAV are reported in Fig. 1.The Facebook Terragraph channel sounders operate inthe IEEE 802.11ad bands [19]. The radios feature TX andRX arrays of 36 × ◦ and 64 beam directions in the azimuth plane,thus with a spacing between each beam of 1 . ◦ . The half-power beamwidth is 2 . ◦ , and the radio maximum effectiveradiated power is 45 dBm. Moreover, the antenna, circuitry,and main board are enclosed in a rugged case, with a smallweight and form factor that makes it possible to deploy themon UAVs. Facebook Terragraph sounders have recently beenused for channel measurement campaigns in the 60GHz bandby several research groups. They are calibrated followingthe procedure in [19] and allow received power, path loss,and Signal to Noise Ratio (SNR) measurements for differenttransmit and receive beam pairs. The standard deviation inpath loss measurements within 1 dB [19]. Some of Facebook Motor 1Motor 2Motor 3Motor 4 Motor 5 Motor 6Intel NUC Facebook TerragraphmmWaveRadioGPS, Compass,Gyro,Accelerometer
Intel NUC DJI M600 Pro Sounder config and readings Ground Controller
Locationreadings
DC-DC step up
MOTOR 1MOTOR 2MOTOR 3MOTOR 4MOTOR 6MOTOR 5
Flight mgmtcommands +_+_+_+_+_+_
Power Module E S C D J I F li g h t C o n t r o ll e r U n i t ( F C U ) M A GG N SS A cc e l . G y r o I/O BoardA3 ProDJI FC J T A G DJI API
DJI Onboard SDKControl Module U S B . DC power inputWi-Fi
TerragraphSounder Module
Eth0 mmWave Radio
60 GHz RF
Eth0 DC in
UAV 1 G r o und W i - F i Flight Control ModuleDJI M600 Pro
UAV 2
Intel NUC mmWaveRadio E t h W i - F i Flight Control Commands
Sounder config and readingsFlight Trajectory
Figure 1: UAV-mounted mmWave channel sounder: Proto-type and hardware schematics. ear Burlington — Middlesex
366 9 12 15 18 21 24 27 3233 403028 1512600
TX RXLegend: m m Figure 2: Deployment scenario.
Terragraph previous use and documentation can be found at[5, 16, 19–21].In our measurement campaign, we run extensive channelsounding experiments on a wide open field with few reflec-tors and scarce multi-path effect as illustrated in Fig. 2. Foreach experiment, the transmitter and receiver UAVs hoverat the same altitude and face each other in full Line of Sight(LOS) conditions. Through our experiment, we consider 3heights, 6 m, 12 m, and 15 m; and a total of 14 distancesbetween transmitter and receiver, namely 6, 9, 12, 15, 18,21, 24, 27, 28, 30, 32, 33, 36, and 40 m, as shown in Fig. 2.The selection of these coordinates has been constrained bythe size of the UAV flying facility. We performed channelmeasurements for channel 2 of the IEEE 802.11ad standard(i.e., the carrier frequency is 60 .
48 GHz, and the bandwidthif 2 .
16 GHz) and 27 altitude-distance pairs, accounting for 3days of flight experiments. Each channel sounding experi-ment consists of scanning the 400 beam pairs, between ± ◦ from the boresight direction (see [19] for specifications),with 15 independent measurements per beam scan, to aver-age small scale fading, and a distance-adaptive transmissiongain. This section describes the data analysis based on the mea-surement campaign outlined in Sec. 2.
The literature on propagation models has proposed severalexperimental path loss laws that can be used to fit the mea-surement results as a function of distance and frequencyparameters. A review can be found in [15]. In this paper, asthe measurements have been taken for a single carrier fre-quency (i.e., 60.48 GHz), we focus on a distance-dependentfit only. A video of the experiments is available at https://youtu.be/Jzwt-tEp98g.
The Close-in free space reference (CI) path loss modelscan be expressed as [15] PL CI ( d , f ) = PL FS , ref ( f ) + n CI log ( d ) + ξ σ , CI . (1)The first term, PL FS , ref ( f ) , is used to model the dependenceon the carrier frequency, and is calculated using Friis’ law forfree space propagation, at the reference distance of 1 m [18]: PL FS , ref ( f ) =
20 log (cid:18) π fc (cid:19) , (2)where c is the speed of light. The second term accounts forthe logarithmic distance-dependent behavior, with n CI thepath loss exponent (PLE), given by the value that best fitsthe measurement data. Finally, ξ σ , CI is a shadow fading termthat in the decibel domain follows a Gaussian distributionwith zero mean and standard deviation σ . The CI modelis widely used for empirical path loss fitting, either in thesingle-slope version of Eq. (1) [2, 15], or with a dual-slopeextension, in which different values of n CI are consideredbefore and after a breakpoint distance [2, 22].Other widespread models belong to the Floating Intercept(FI) family, in which the term based on Friis’ law is replacedby a generic value, determined based on the best fit on thedata. A notable example is the Alpha-Beta-Gamma (ABG)model, given by [15] PL ABG ( d , f ) = β + γ log ( f ) + α log ( d ) + ξ σ , ABG , (3)where ξ σ , ABG is a Gaussian shadow fading with zero meanand standard deviation σ , as in the CI model, and β , γ , and α are fit on the data. With respect to the CI model in Eq. (1), theterm α is equivalent to the path loss exponent n CI . Noticethat, as the measurement campaign of this paper is basedon a single frequency, we cannot compute both β and γ .Therefore, in the following, we will consider a simplifiedversion, with two fit parameters, given by PL F I ( d ) = PL F I + n F I log ( d ) + ξ σ , F I , (4)where PL F I = β + γ log ( f ) , n F I = α , and ξ σ , F I = ξ σ , ABG .For both models, the fit parameters (i.e., n CI for CI, and n F I and PL F I for FI) are computed as the slope (and the intercept,for FI) of a linear fit on the logarithm of the distance, and thestandard deviation of the shadow fading is computed as theroot mean square error on the fit [3, 15].
We first discuss whether a CI fit is representative of the A2Apath loss measurements, or whether a FI fit is preferable.Figure 3 compares the two methods, considering data for allthe values for the height (i.e., h ∈ [ , , ] m) and distance d from 6 to 40 m. Both curves share the same trend, with theFI fit slightly steeper, but with a difference of at most 0 . d = . d =
40 m. Table 3 reports the fitparameters for both, showing that the FI identifies a fit for its
UAV-to-UAV distance [m] P a t h l o ss [ d B ] Meas., h = m Meas., h = m Meas., h = mCI fit, all heights FI fit, all heights Figure 3: Comparison of CI and FI fits.
Intercept [dB] Path loss exponent σ [dB]CI fit 68.08 2.25 3.56FI fit 67.03 2.33 3.52 Table 1: Comparison of the parameters for CI and FI fits, con-sidering all heights h ∈ [ , , ] m. floating intercept at a 67 .
03 dB, which is 1 .
05 dB smaller thanthe free space path loss in the same conditions, and a pathloss exponent of 2 .
33, against the value of 2 .
25 for the CI fit.The shadowing standard deviation σ is 0 .
04 dB smaller forthe FI fit, which is almost two orders of magnitude smallerthan the actual value of σ (i.e., 3 .
52 dB for FI and 3 .
56 dB forCI). Following standard practices in the literature [18], weconclude that both models provide a suitable representationof the path loss in a A2A 60 GHz link. Given this, in theremainder of the paper, we will consider the CI fit as baseline,as it is simpler than the FI fit (i.e., one fit parameter instead oftwo), and is based on the fundamental principles of wirelesspropagation through the Friis-based intercept [18].
As the deployment height of UAV networks is subject toapplication scenario, regulations, and performance require-ments [3], it is important to characterize the channel behav-ior for a wide range of deployment heights. Therefore, inFig. 4 we investigate the impact of the height on the param-eters of a CI fit. For this, we consider separate CI fits forthe three values of the UAVs height at which the measure-ments were collected, i.e., h =
6, 12, and 15 m, and comparethem with a CI fit that does not distinguish between differ-ent heights. As can be seen, the four curves in Fig. 4 share
All heights h = h =
12 m h =
15 mPLE n CI σ [dB] Table 2: Parameters for the CI fit for different heights.
10 20 30 408090100110
UAV-to-UAV distance [m] P a t h l o ss [ d B ] CI fit, h = m CI fit, h = m CI fit, h = mCI fit, all heights Meas., h = mMeas., h = mMeas., h = m Figure 4: Comparison of different CI fits considering mea-surements at different heights, and a CI fit that combinesall the measurements.
UAV-to-UAV distance [m] P a t h l o ss [ d B ] UAV CI fit 3GPP UMi model 3GPP UMa model3GPP RMa model 3GPP InOo model Free space pathloss
Figure 5: Comparison of CI fit, 3GPP and free space path losschannel models. the same trend, with differences of less than 1 dB betweenthe curves for h = h =
15 m in the worst case.The path loss exponent n CI and the standard deviation areshown in Table 2. The value of n CI is very similar for thefour fits, showing that the height does not a significant im-pact once the UAVs are in flight. The standard deviation σ shows a higher variability, also with respect to the FI fit pre-viously described, but this can be traced back to the fewermeasurements considered for the fit at each single heightvalue. As discussed in Sec. 1, aerial communications are affectedby a different channel than terrestrial networks, where bothendpoints are on the ground and present less erratic mobilitypatterns. An aerial link is indeed characterized by a partic-ularly strong LOS link, while reflections provide limited orno contribution, as scatterers are at a larger distance than inellular or indoor scenarios. Furthermore, the micro-mobilityof the UAVs, given by the fluctuations caused by the UAVcontrol loop and/or wind conditions, may degrade the linkquality [7].Figure 5 shows that the measurement-based fit we proposein this paper reflects these phenomena into a higher pathloss exponent with respect to free space path loss and of3rd Generation Partnership Project (3GPP) channel models.We consider the equations for LOS propagation loss fromthe 3GPP channel model for frequencies between 0.5 and100 GHz [2], with an additional distance-dependent oxygenabsorption loss factor for the 60 GHz band. We compare thepath loss of different 3GPP scenarios, i.e., for Rural Macro(RMa), Urban Macro (UMa), Urban Micro (UMi) and IndoorOpen Office (InOo) deployments. The free space path lossfollows Friis’ law [15]. The path loss curves in Fig. 5 confirmthat the propagation of 60 GHz signals experiences a higherloss in an aerial link, as it has a path loss exponent n CI = .
25 which is larger than the worse LOS exponent in 3GPPscenarios (i.e., 2.1 for UMi). This confirms that the analysisand simulation of aerial networks cannot be based on channelmodels developed for terrestrial applications , and motivates thedevelopment of the experimental channel model proposedin this paper.
UAV-to-UAV distance [m] P a t h l o ss [ d B ] Figure 7: FI fit for the path loss of first 9 beam pairs. mmWave links will rely on beamforming techniques to in-crease the link budget and compensate for the increased pathloss. In this sense, a proper beam alignment makes it possi-ble to select the transmit and receive beams that yield thehighest beamforming gain. Fast and prompt beam tracking,however, can be challenging in aerial links, where the nodesare highly mobile [5]. Therefore, we present in the followingparagraph the path loss curves that fit measurements forbeam pairs that are not perfectly aligned, following a similar
10 20 30 4090100110120
UAV-to-UAV distance [m] P a t h l o ss [ d B ] FI fit, second best beam pairMeas., second best beam pairCI fit, best beam pair (a) FI fit for the path loss of the second-best beam pair.
10 20 30 4090100110120
UAV-to-UAV distance [m] P a t h l o ss [ d B ] FI fit, third best beam pairMeas., third best beam pairCI fit, best beam pair (b) FI fit for the path loss of the third-best beam pair.
10 20 30 4090100110120
UAV-to-UAV distance [m] P a t h l o ss [ d B ] FI fit, ninth best beam pairMeas., ninth best beam pairCI fit, best beam pair (c) FI fit for the path loss of the ninth-best beam pair.Figure 6: Measurements vs. FI fits for the path loss of additional beam pairs.
Best beam pair 2nd best 3rd best 4th best 5th best 6th best 7th best 8th best 9th bestPLE n σ [dB] 3.56 3.78 4.85 4.61 4.01 5.76 5.80 5.38 4.82Displacement ∆ [deg] 0 1.87 2.59 2.70 3.47 3.42 4.20 3.89 4.20 Table 3: Parameters for the fit for different best beam pairs. pproach discussed in [14] for terahertz links. Notice thatthese results do not represent the actual path loss (whichis given only by the data for the best beam pair), but are apractical way to model the loss in beamforming gain due tomisalignment.For this analysis, we exploit the beam scanning capabilitiesof the Terragraph channel sounders [19]. For each distanceand height at which measurements are taken, the beam pairthat yields the lowest path loss is considered for the CI fitpreviously described. We then analyze the remaining beampairs, and, for each distance-height point, select the second-through-ninth best beam pairs in terms of path loss. Finally,we apply an FI fit for each of these, by aggregating the datafor different heights. For this analysis, the FI fit is preferredover the CI fit, as its intercept term PL F I allows the modelingof the additional loss (with respect to the free space pathloss) introduced by beam misalignment.As an example, Fig. 6 reports the measurements and fittedpath loss curves for the second, third and ninth-best beampairs, and the path loss curve for the best beam pair. Weobserve that the distance between measured points and thebest beam pair fit line progressively increases, as the mis-alignment between the non-optimal beam pairs decreasesthe beamforming gain. Figure 7, instead, directly comparespath loss curves, which exhibit the same trend, as the phys-ical path loss between the two nodes does not change, buthave an increasing loss factor that models the reduction inbeamforming gain.To further underline this, we list in Table 3 the path lossexponents, which are comparable for the different beampairs, and the intercept values, which increase by up to11 dB from the best to the ninth-best beam pair. We alsoprovide a displacement metric, to model the angular er-ror that can lead to such loss in beamforming gain. No-tably, for the i -th-best beam pair, i ≥
2, we have ∆ i = E (cid:2) | θ tx , best − θ tx , i | + | θ rx , best − θ rx , i | (cid:3) , with E the averageoperator, and θ j , best and θ j , i the angles of the best beam, andof the beam associated to the i -th beam pair, for j ∈ { tx , rx } ,respectively. As discussed in Sec. 2, the beams of the Terra-graph sounders are spaced by 1 . ◦ . Therefore, ∆ = . ◦ ,for the second beam pair, suggests that, on average, the beam-forming gain loss is due to a misalignment of 1.33 beam pairs,either at the receiver or the transmitter. We notice that ∆ saturates after the fifth-best beam pair, and that the intercept PL F I has a limited difference, showing that the misalignmentis likely due to the selection of equivalent beam pairs.
In this paper, we described an extensive measurement cam-paign for the characterization of the A2A path loss at 60 GHz, using the Facebook Terragraph radios configured as chan-nel sounders. The data analysis led us to conclude that thepropagation loss of an aerial link, for heights between 6 and15 m, is well represented by the equation: PL CI ( d ) = . + . ( d ) + ξ σ , CI , (5)with ξ σ , CI a Gaussian random variable with standard de-viation σ = .
56 dB. Moreover, we compared this fit withestablished propagation models for mmWaves, and analyzedthe impact of a sub-optimal beam selection on the link per-formance. For this case, the combined path loss and gainreduction can be computed using Eq. (4) with the parametersfrom Table 3.As part of our future work, we will extend the validationof the path loss curves by performing measurements in dif-ferent scenarios and with different sounders, and analyze theimpact of beam misalignment with different beamformingconfigurations, as well as of the Doppler effect in highly mo-bile conditions. Moreover, we will characterize air-to-ground(A2G) mmWave channels and extend the height range forthe A2A measurements.
ACKNOWLEDGEMENTS
This work was supported in part by the US National ScienceFoundation under Grant CNS-1618727 and in part by theUS Office of Naval Research under Grants N00014-19-1-2409and N00014-20-1-2132.
REFERENCES [1] 3GPP. 2018.
NR; Overall description; Stage-2 . Technical Specification(TS) 38.300. Version 15.0.0.[2] 3GPP. 2018.
Study on channel model for frequencies from 0.5 to 100 GHz .Technical Report (TR) 38.901. Version 15.0.0.[3] R. Amorim, H. Nguyen, P. Mogensen, I. Z. KovÃącs, J. Wigard, and T. B.SÃÿrensen. 2017. Radio Channel Modeling for UAV CommunicationOver Cellular Networks.
IEEE Wireless Communications Letters
6, 4(Aug 2017), 514–517. https://doi.org/10.1109/LWC.2017.2710045[4] L. Bertizzolo, S. D’oro, L. Ferranti, L. Bonati, E. Demirors, Z. Guan,T. Melodia, and S. Pudlewski. 2020. SwarmControl: An AutomatedDistributed Control Framework for Self-Optimizing Drone Networks.In
Proc. of IEEE Conf. on Computer Communications (INFOCOM) .[5] L. Bertizzolo, M. Polese, L. Bonati, A. Gosain, M. Zorzi, and T. Melodia.2019. mmBAC: Location- aided mmWave Backhaul Management forUAV-based Aerial Cells. In
Proc. of 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems (mmNets) . Los Cabos, Mexico.[6] L. Bertizzolo, T. X. Tran, B. Amento, B. Balasubramanian, R. Jana, H.Purdy, Y. Zhou, and T. Melodia. 2020. Live and Let Live: Flying UAVsWithout Affecting Terrestrial UEs. In
Proc. of the 21st Intl. Workshopon Mobile Comput. Systems and Appl. (HotMobile) . Austin, TX, USA.[7] M. T. Dabiri, H. Safi, S. Parsaeefard, and W. Saad. 2020. AnalyticalChannel Models for Millimeter Wave UAV Networks Under HoveringFluctuations.
IEEE Trans. Wireless Commun.
19, 4 (Feb 2020), 2868–2883.[8] Terragraph Facebook. 2019. https://terragraph.com/product.[9] M. Gapeyenko, V. Petrov, D. Moltchanov, S. Andreev, N. Himayat,and Y. Koucheryavy. 2018. Flexible and Reliable UAV-Assisted Back-haul Operation in 5G mmWave Cellular Networks.
IEEE J. Sel. Areasommun.
36, 11 (Nov 2018), 2486–2496.[10] W. Khawaja, O. Ozdemir, and I. Guvenc. 2017. UAV Air-to-GroundChannel Characterization for mmWave Systems. In
Proc. of IEEE VTC-Fall . Toronto, ON, Canada.[11] W. Khawaja, O. Ozdemir, and I. Guvenc. 2018. Temporal and SpatialCharacteristics of mmWave Propagation Channels for UAVs. In
Proc.of IEEE Global Symp. on Millimeter Waves (GSMM) . Boulder, CO, USA.[12] A. A. Khuwaja, Y. Chen, N. Zhao, M. Alouini, and P. Dobbins. 2018. ASurvey of Channel Modeling for UAV Communications.
IEEE Commun.Surveys Tuts.
20, 4 (Fourth quarter 2018), 2804–2821.[13] M. Polese, M. Giordani, T. Zugno, A. Roy, S. Goyal, D. Castor, and M.Zorzi. 2020. Integrated Access and Backhaul in 5G mmWave Networks:Potential and Challenges.
IEEE Commun. Mag.
58, 3 (Mar 2020), 62–68.[14] S. Priebe, M. Jacob, and T. KÃijrner. 2012. Affection of THz indoorcommunication links by antenna misalignment. In
Proc. of 6th EuropeanConf. on Antennas and Propagation (EUCAP) . Prague, Czech Republic,483–487. https://doi.org/10.1109/EuCAP.2012.6206026[15] T. S. Rappaport, Y. Xing, G. R. MacCartney, A. F. Molisch, E. Mellios,and J. Zhang. 2017. Overview of Millimeter Wave Communicationsfor Fifth-Generation (5G) Wireless NetworksâĂŤWith a Focus onPropagation Models.
IEEE Trans. Antennas Propag.
65, 12 (Dec 2017),6213–6230. [16] Maryam Eslami Rasekh, Zhinus Marzi, Yanzi Zhu, Upamanyu Madhow,and Haitao Zheng. 2017. Noncoherent mmWave path tracking. In
Proc. of the 18th Intl. Workshop on Mobile Computing Systems andApplications . Sonoma, CA, USA, 13–18.[17] V. Semkin, J. Haarla, T. Pairon, C. Slezak, S. Rangan, V. Viikari, and C.Oestges. 2020. Analyzing Radar Cross Section Signatures of DiverseDrone Models at mmWave Frequencies.
IEEE Access