Neural Network-Based Ranging with LTE Channel Impulse Response for Localization in Indoor Environments
Halim Lee, Ali A. Abdallah, Jongmin Park, Jiwon Seo, Zaher M. Kassas
aa r X i v : . [ ee ss . SP ] S e p ∼
16, 2020; BEXCO, Busan, Korea
Neural Network-Based Ranging with LTE Channel Impulse Response forLocalization in Indoor Environments
Halim Lee , Ali A. Abdallah , Jongmin Park , Jiwon Seo ∗ , and Zaher M. Kassas , School of Integrated Technology, Yonsei University,Incheon 21983, Korea (halim.lee, jm97, [email protected]) Department of Electrical Engineering and Computer Science, University of California, Irvine, Department of Mechanical and Aerospace Engineering, University of California, Irvine,California 92697, USA ([email protected], [email protected]) ∗ Corresponding author
Abstract:
A neural network (NN)-based approach for indoor localization via cellular long-term evolution (LTE) signalsis proposed. The approach estimates, from the channel impulse response (CIR), the range between an LTE eNodeB and areceiver. A software-defined radio (SDR) extracts the CIR, which is fed to a long short-term memory model (LSTM) re-current neural network (RNN) to estimate the range. Experimental results are presented comparing the proposed approachagainst a baseline RNN without LSTM. The results show a receiver navigating for 100 m in an indoor environment, whilereceiving signals from one LTE eNodeB. The ranging root-mean squared error (RMSE) and ranging maximum error alongthe receiver’s trajectory were reduced from 13.11 m and 55.68 m, respectively, in the baseline RNN to 9.02 m and 27.40m, respectively, with the proposed RNN-LSTM.
Keywords: long-term evolution (LTE), indoor localization, indoor navigation, recurrent neural network (RNN), longshort-term memory (LSTM)
1. INTRODUCTION
Location-based services (LBS) have become an es-sential part of our lives [1]. LBS depend on navigationtechnologies, such as global navigation satellite system(GNSS) [2, 3] and enhanced long-range navigation (eLo-ran) [4, 5]. LBS can also exploit other radio signals inthe environment, such as Wi-Fi [6–8] and cellular signals[9–12].In outdoor environments, GNSS provide an accept-able localization performance. A receiver’s position canbe estimated to within a few meters utilizing pseudor-ange measurements [13, 14], while decimeter-level ac-curacy is achievable with carrier phase measurements[15]. However, GNSS are vulnerable to radio frequencyinterference [16–18] and atmospheric changes, such asionosphere anomalies [19, 20]. In indoor environments,GNSS signals get severely attenuated, making them prac-tically unusable. For indoor environments, significantattention has been devoted to localization with WiFi[21, 22], ultra-wide band (UWB) [23, 24], and radio-frequency identification (RFID) [25, 26].Cellular signals, particularly long-term evolution(LTE) signals, have shown tremendous promise in cir-cumventing the limitations of GNSS signals in both in-door and outdoor environments [27]. Cellular signals canbe exploited to produce a standalone navigation solutionor can be coupled with other sensors (e.g., lidar, iner-tial measurement unit, etc.). In outdoor environments,recent work has shown that cellular signals could yieldmeter-level and even lane-level accurate localization onground vehicles [28–32] and sub-meter-level accurate lo-calization on aerial vehicles [33,34]. Meter-level accuratelocalization has been recently reported indoors [35–39]. What makes cellular LTE signals especially attractive istheir ubiquity, high power, and geometric diversity.Two types of ranging method are commonly used forLTE-based localization: signal strength-based and time-of-arrival (TOA)-based. Among the two methods, signal-strength-based ranging has the advantage of low com-plexity. Accordingly, many researchers have studiedreceived signal strength indicator (RSSI)-based rangingmethods [40, 41].RSSI is the average of total received power observedin orthogonal frequency-division multiplexing (OFDM)reference symbols [42]. Although it is possible to per-form ranging using a channel model and RSSI, the ac-curacy is rather low because of signal reflection and at-tenuation caused by obstacles. Particularly in indoor en-vironments, because of rapid time-varying channels andthe presence of many obstacles, it is difficult to model thechannel accurately in real-time, reducing the accuracy ofRSSI-based ranging.Unlike RSSI, the channel frequency response (CFR)provides detailed information about the channel. LTE re-ceivers estimate CFR from the cell-specific reference sig-nal (CRS) of the LTE physical layer. The CFR providesinformation about the channel experienced by each sym-bol in the frequency domain.Fingerprinting-based localization using channel infor-mation extracted from an LTE downlink signal has beenpreviously studied [43–45]. In [43], a feed-forward three-layer neural-network-based LTE fingerprinting methodwas suggested. Eleven channel parameters were usedas input to the neural network, which were extractedfrom the channel impulse response (CIR). The CIR isthe inverse Fourier transform of the CFR. In other stud-ies [44, 45], channel state information (CSI)-based fin-erprinting methods were suggested. CSI is a technicalterm for the channel response used in IEEE 802.11 a/g/nstandards [46]. In [44], a CSI descriptor-based nearest-neighbor fingerprinting algorithm was presented. TheCSI descriptors are composed of elements expressing CSIcharacteristics, such as the mean, standard deviation, andFano factors of CSI. In addition, a two-stage cascadedneural network was introduced [45].However, fingerprinting-based localization requires alarge number of fingerprinting maps. Considering the dif-ficulty to survey large areas to obtain and maintain the fin-gerprinting maps, localization based on range measure-ments from a user to nearby eNodeB is more practical.This paper proposes a recurrent neural network (RNN)-based ranging method using CIR. The CIR was obtainedfrom real LTE signals, and the range between the userequipment (UE) and the LTE eNodeB was estimated. Along short-term memory model (LSTM) network was de-signed to extract the range from the magnitude of CIR.Experimental results are presented comparing the pro-posed approach against a baseline recurrent neural net-work (RNN) without LSTM. The results show a receivernavigating for 109 m in an indoor environment, while re-ceiving signals from one LTE eNodeB. The ranging root-mean squared error (RMSE) and ranging maximum erroralong the receiver’s trajectory were reduced from 13.11m and 55.68 m, respectively, in the baseline NN to 8.01m and 31.57 m, respectively, with the proposed RNN-LSTM.The remainder of this paper is organized as fol-lows. Section 2 describes the proposed RNN-LSTM-based ranging approach. Section 3 presents experimen-tal results in an indoor environment. Section 4 presentsconcluding remarks.
2. PROPOSED METHOD
This section presents: (i) CIR estimation of the re-ceived LTE signals and (ii) proposed RNN-LSTM-basedmodel to estimate ranges from the LTE CIRs.
The LTE system uses orthogonal frequency divisionmultiplexing (OFDM) as a modulation technique with aframe duration of 10 ms and subcarrier spacing ∆ f =15 kHz. In the time-domain, an LTE frame is dividedinto ten subframes, where each subframe is divided intotwo slots and each slot consists of seven OFDM symbols.In the-frequency domain, the LTE bandwidth is scalablefrom 1.4 MHz to 20 MHz with a different number of totaland used subcarriers in each configuration.CIR estimation can be performed using any of the sev-eral LTE reference signals such as: (i) primary synchro-nization signal (PSS), (ii) secondary synchronization sig-nal (SSS), and (iii) CRS. The PSS and SSS are transmit-ted to provide the frame start time and the cell ID of theLTE base station, also known as the evolved Node B (eN-odeB); however, the PSS and SSS have a fixed bandwidthof 0.93 MHz regardless of the LTE downlink bandwidth. Subframe1 2 3 4 5 6 7 8 90 S ub c a rr i e r s u s e d ( N r ) Slot CRSSSSPSS T f = 10 ms)0.5 ms Time F r e qu e n c y PRS
Fig. 1 LTE frame structure [37].On the other hand, the CRS bandwidth is the same as thetransmission bandwidth (i.e., up to 20 MHz). This makesthe CRS more attractive for range measurements, espe-cially in multipath environments. The LTE frame struc-ture is shown in Fig. 1.This paper adopts the carrier phase-based software de-fined receiver (SDR) proposed in [47] in which the CIR isestimated by tracking the CRS. In this receiver, there aretwo stages: (i) acquisition stage and (ii) tracking stage. Inthe acquisition stage, the received LTE baseband signalis correlated with all possible locally-generated PSS andSSS sequences to produce a coarse estimate of the framestart time, which is used to control the fast Fourier trans-form (FFT) window timing. The LTE guard band, alsoknown as cyclic prefix (CP), elements are removed andan FFT is taken to convert the signal into the LTE framestructure. Then, the CIR is estimated using the estimationof signal parameters via rotational invariance techniques(ESPRIT) and used to refine the time-of-arrival (TOA).The phase difference between CFRs estimated from twodistinct CRS symbols is used to provide a coarse estimateof Doppler frequency, ˆ f D .In the tracking stage, a phase-locked loop (PLL) is im-plemented to track the phase of the CRS signal. The car-rier phase discriminator can be defined as the phase of theintegrated CFRs over the entire subcarrier [48]. Then, asecond-order loop filter at the output of the discriminatorcan be used to estimate the rate of change of the carrierphase error, π ˆ f D , expressed in rad/s. Finally, the TOAestimate, ˆ e τ , is updated according to ˆ e τ ←− ˆ e τ − T f T s v PLL , (1)where T f = 10 ms, T s is the sampling time, and v PLL isthe output of the PLL.
Since the CIR is a temporal sequence, an RNN, whichhas advantages in solving sequential problems, was de-signed. The RNN was long short-term memory (LSTM)-based and is depicted in Fig. 2.The proposed RNN-LSTM consists of one embedding mbedding Embedding Embedding EmbeddingTime-aligned channel impulse responses LSTMLSTMLSTMLSTM DenseDenseEstimated Range
Fig. 2 Proposed RNN-LSTM structure. The time-aligned CIR and estimated range are the input and outputof the proposed RNN-LSTM, respectively. The RNN-LSTM consists of embedding, LSTM, and dense layers.layer, one LSTM layer, and two dense layers. The out-put dimension of the embedding layer is set to 128. Thenumber of the hidden units of the LSTM layer are set to128. Further, the output dimensions of the dense layersare set to 128 and 1 for the first and second dense layers,respectively. The rectified linear unit (ReLU) is used asthe activation function of dense layers. The output of thelast activation function is the estimated range.
3. EXPERIMENTAL RESULTS
To validate the proposed approach, an experiment wasconducted in an indoor environment: Winston ChungHall building at the University of California, Riverside,USA. Two LTE receivers were used: (i) a rover receiverwhich navigates indoors and (ii) a base receiver which isplaced on the roof of the building. The base receiver hasaccess to GPS and is used to estimate the clock biases ofthe LTE eNodeB. The estimated eNodeB’s clock biaseswere removed from the rover’s measurements; hence, therover’s measurement errors are mainly affected by themultipath. The base receiver was equipped with a single-channel National Instruments (NI) universal software ra-dio peripheral (USRP)-2920 to simultaneously down-mixand synchronously sample LTE signals at 10 Msps. Therover’s receiver was equipped with a dual-channel USRP-2954R; however, one channel is exploited here to samplethe LTE signals at the sampling rate of 20 Msps. Bothreceivers were equipped with a consumer-grade cellularomnidirectional antenna to collect LTE data at the carrierfrequency of 2145 MHz, which corresponds to the U.S.cellular provider T-Mobile.The collected data were processed in a post-processingfashion using the receiver discussed in Subsection 2.1,where the eNodeB (physical cell ID: 383) was tracked.Throughout the experiment, a smart phone was used torecord the location of pre-placed tags at known locationon the ground, which was later processed and used as aground truth. Fig. 3 shows the environmental layout andthe hardware setup for both the rover and the base re-
USRP-2954R USRP-2920 (c)(a) (b)
Fig. 3 Environmental layout and experimental setup. (a)shows the rover’s hardware setup, (b) shows the base’shardware setup, and (c) shows the environmental layout,eNodeB location, base location, and rover’s trajectory.ceivers.
Throughout the experiment, the receiver travelled adistance of 109 m in 50 seconds. The CRS-based CIRwas estimated for each received LTE frame, which pro-vides 5000 samples that were divided as the training, val-idation, and test data, which were , , and ofthe total number of samples, respectively. For the train-ing process, an AdamOptimizer with the learning rate of − was used to minimize the mean squared error of theestimated ranges. The LSTM network was implementedon TensorFlow. The learning was conducted using thetraining and validation data, and the performance wasevaluated using the test data. The training and valida-tion losses with respect to the training epoch are shownin Fig. 4. The training and validation losses are definedas the root-mean-squared error (RMSE) of the differencebetween the estimated range ˆ r i and the true range r i , asshown in Eq. (2). LOSS = vuut n n X i =1 ( ˆ r i − r i ) , (2)where n is the total number of training or validation sam-ples. The performance of the proposed RNN-LSTM is com-pared with that of a baseline RNN. The baseline RNN
10 20 30 40 50 epoch R M SE Trainingvalidation
Fig. 4 Training and validation losses with respect to thetraining epoch.that we designed for the purpose of comparison consistsof two dense layers with the ReLU function as the activa-tion function. The output dimensions of the dense layerswere set to 128 and 1 for the first and second dense lay-ers, respectively. The baseline RNN was trained with anAdamOptimizer with the learning rate of − for 300epochs. The only difference between the design of theproposed network and the baseline RNN is the existenceof the embedding and LSTM layers for sequential pro-cessing. The only difference between the proposed net-work and the baseline RNN is the existence of the em-bedding and LSTM layers for sequential processing. Theoutput dimensions, activation function of dense layers,learning rate, and optimizer remain the same for both net-works.Fig. 5 (a) shows the cumulative distribution func-tions (CDFs) of the ranging errors for the proposed RNN-LSTM and the baseline RNN. Fig. 5 (b) shows the rang-ing errors of the proposed RNN-LSTM and the baselineRNN for each test sample. Further, Table 1 compares theranging performance of the two networks. The proposedRNN-LSTM exhibited the ranging RMSE of 9.02 m, out-performing the 13.11 m RMSE of the baseline RNN.After suspecting that this performance could be causedby overfitting, we tried several approaches that are knownto resolve the overfitting problem. First, the learning ratedecay [49] was tried to help the optimization process oftraining. A stochastic gradient descent (SGD) optimizerwith an initial learning rate of 0.1, a decay step of 1000,and a decay rate of 0.96 was used. However, the rangingRMSE of the optimal result obtained by using the learn-ing rate decay was 10.78 m, which is worse than thatof the proposed RNN-LSTM without the application ofthe learning rate decay. Second, we applied the dropout,which is a technique to prevent overfitting by randomlydrop units from the neural network during training [50].The ranging RMSE obtained by adding a dropout func-tion between the LSTM layer and the dense layer withthe dropout rate of 0.2 was 9.76 m, which is not betterthan the performance of the proposed RNN-LSTM with-out the application of the dropout. Finally, we tried morecomplex layers and then observed overfitting. For exam-ple, when the neural network consisted of one embeddinglayer, one gated recurrent unit (GRU) layer, one LSTMlayer, and three dense layers was tested and the learning rate decay was applied, all estimated ranges convergedto the same value. The converged value was 332.82 m,which is very close to the mean of the true ranges in thetraining data (332.67 m). Based on these observations,we concluded that the proposed RNN-LSTM does notsuffer from overfitting. Since the data used in this studywas collected in a very challenging multipath environ-ment, the data has low generality characteristics. Testing sample R ang i ng e rr o r [ m ] Ranging error [m] CD F (b)(a) Fig. 5 (a) CDFs of the ranging errors of the proposedRNN-LSTM and baseline RNN. (b) Ranging errors ofboth approaches for each test sample out of the 1000 testsamples.Table 1 Ranging Performance Comparison
Performance Measure [m] BaselineRNN ProposedRNN-LSTM
RMSE 13.11 9.02Standard deviation 9.17 5.40Maximum error 55.68 27.40
4. CONCLUSION
An RNN-LSTM-based approach to estimate the rangebetween a UE and an LTE eNodeB using the CIR ex-tracted from the CRS signal was developed. An RNNutilizing timely-synchronized magnitudes of CIR was de-signed. The proposed approach was validated experimen-tally, where LTE signals were collected in an indoor en-vironment over 109 m in 50 seconds. The proposed ap-proach reduced the ranging RMSE by 68.8% comparedto the performance achieved by a baseline RNN.
ACKNOWLEDGEMENT
The authors would like to thank Joe Khalife, KimiaShamaei, and Mahdi Maaref for their help in data collec-ion. This research was supported by the Ministry of Sci-ence and ICT (MSIT), Korea, under the High-PotentialIndividuals Global Training Program (2020-0-01531) su-pervised by the Institute for Information & Communica-tions Technology Planning & Evaluation (IITP) and alsosupported by the HPC Support Project funded by MSITand the National IT Industry Promotion Agency (NIPA)of Korea. This work was partially performed under the fi-nancial assistance award 70NANB17H192 from U.S. De-partment of Commerce, National Institute of Standardsand Technology (NIST).
REFERENCES [1] G. World, “LBS marketworth $39.87 billion by 2019,”http://gpsworld.com/lbs-market-worth-39-87-billion-by-2019,April 2014.[2] N. Gogoi, A. Minetto, and F. Dovis, “On the cooper-ative ranging between android smartphones sharingraw GNSS measurements,” in
Proceedings of IEEEVehicular Technology Conference , 2019, pp. 1–5.[3] H. Sun, M. Dong, B. Chu, M. Ao, C. Chen, andS. Gu, “Multi-level high precision LBS architec-ture based on GNSS CORS network, a case studyof HNCORS,”
IEEE Access , vol. 7, pp. 185 042–185 054, 2019.[4] S. Lo, B. P. amd T. Hardy, and P. Enge, “Improv-ing Loran coverage with low power transmitters,”
NAVIGATION, Journal of the Institute of Naviga-tion , vol. 63, no. 1, p. 23, January 2010.[5] P.-W. Son, J. Rhee, J. Hwang, and J. Seo, “Univer-sal kriging for Loran ASF map generation,”
IEEETransactions on Aerospace and Electronic Systems ,vol. 55, no. 4, pp. 1828–1842, 2019.[6] R. Faragher, C. Sarno, and M. Newman, “Op-portunistic radio SLAM for indoor navigationusing smartphone sensors,” in
Proceedings ofIEEE/ION Position Location and Navigation Sym-posium , April 2012, pp. 120–128.[7] C. Yang and H. Shao, “WiFi-based indoor posi-tioning,”
IEEE Communications Magazine , vol. 53,no. 3, pp. 150–157, March 2015.[8] J. Khalife, Z. Kassas, and S. Saab, “Indoor localiza-tion based on floor plans and power maps: Non-lineof sight to virtual line of sight,” in
Proceedings ofION GNSS Conference , September 2015, pp. 2291–2300.[9] C. Gentner, E. Munoz, M. Khider, E. Staudinger,S. Sand, and A. Dammann, “Particle filter based po-sitioning with 3GPP-LTE in indoor environments,”in
Proceedings of IEEE/ION Position, Location,and Navigation Symposium , April 2012, pp. 301–308.[10] K. Shamaei, J. Khalife, and Z. Kassas, “ExploitingLTE signals for navigation: Theory to implementa-tion,”
IEEE Transactions on Wireless Communica-tions , vol. 17, no. 4, pp. 2173–2189, April 2018.[11] A. Abdallah, S. Saab, and Z. Kassas, “A machine learning approach for localization in cellular envi-ronments,” in
Proceedings of IEEE/ION Position,Location, and Navigation Symposium , April 2018,pp. 1223–1227.[12] J. del Peral-Rosado, R. Raulefs, J. L´opez-Salcedo,and G. Seco-Granados, “Survey of cellular mobileradio localization methods: From 1G to 5G,”
IEEECommunications Surveys Tutorials , vol. 20, no. 2,pp. 1124–1148, 2018.[13] D. De Lorenzo, S. Lo, J. Seo, Y. Chen, and P. Enge,“The WAAS/L5 signal for robust time transfer:Adaptive beamsteering antennas for satellite timesynchronization,” in
Proceedings of ION GNSS ,2010.[14] D. Yoon, C. Kee, J. Seo, and B. Park, “Position ac-curacy improvement by implementing the DGNSS-CP algorithm in smartphones,”
Sensors , vol. 16,no. 6, p. 910, 2016.[15] K. Pesyna, Z. Kassas, R. Heath, and T. Humphreys,“A phase-reconstruction technique for low-powercentimeter-accurate mobile positioning,”
IEEETransactions on Signal Processing , vol. 62, no. 10,pp. 2595–2610, May 2014.[16] E. Kim and J. Seo, “SFOL pulse: A high accuracyDME pulse for alternative aircraft position and nav-igation,”
Sensors , vol. 17, no. 10, p. 2183, 2017.[17] K. Park, D. Lee, and J. Seo, “Dual-polarized GPSantenna array algorithm to adaptively mitigate alarge number of interference signals,”
AerospaceScience and Technology , vol. 78, pp. 387–396,2018.[18] P.-W. Son, J. Rhee, and J. Seo, “Novel multichain-based Loran positioning algorithm for resilient nav-igation,”
IEEE Transactions on Aerospace andElectronic Systems , vol. 54, no. 2, pp. 666–679,2018.[19] J. Seo and T. Walter, “Future dual-frequency GPSnavigation system for intelligent air transporta-tion under strong ionospheric scintillation,”
IEEETransactions on Intelligent Transportation Systems ,vol. 15, no. 5, pp. 2224–2236, 2014.[20] J. Lee, Y. Morton, J. Lee, H. Moon, and J. Seo,“Monitoring and mitigation of ionospheric anoma-lies for GNSS-based safety critical systems,”
IEEESignal Processing Magazine , vol. 34, no. 5, pp. 96–110, 2017.[21] R. Faragher and R. Harle, “Towards an efficient, in-telligent, opportunistic smartphone indoor position-ing system,”
NAVIGATION, Journal of the Instituteof Navigation , vol. 62, no. 1, pp. 55–72, 2015.[22] Z. Zhang, S. He, Y. Shu, and Z. Shi, “A self-evolving WiFi-based indoor navigation system us-ing smartphones,”
IEEE Transactions on MobileComputing , vol. 19, no. 8, pp. 1760–1774, 2020.[23] O. Bialer, D. Raphaeli, and A. Weiss, “Efficienttime of arrival estimation algorithm achieving max-imum likelihood performance in dense multipath,”
IEEE Transactions on Signal Processing , vol. 60,o. 3, pp. 1241–1252, 2012.[24] Y. Xu, Y. Shmaliy, Y. Li, and X. Chen, “UWB-based indoor human localization with time-delayeddata using EFIR filtering,”
IEEE Access , vol. 5, pp.16 676–16 683, 2017.[25] C. Yao and W. Hsia, “An indoor positioning sys-tem based on the dual-channel passive RFID tech-nology,”
IEEE Transactions on Instrumentation andMeasurement , vol. 18, no. 11, pp. 4654–4663, 2018.[26] Y. Zeng, X. Chen, R. Li, and H. Tan, “UHF RFIDindoor positioning system with phase interferencemodel based on double tag array,”
IEEE Access ,vol. 7, pp. 76 768–76 778, 2019.[27] Z. Kassas, J. Khalife, K. Shamaei, and J. Morales,“I hear, therefore I know where I am: Compensatingfor GNSS limitations with cellular signals,”
IEEESignal Processing Magazine , pp. 111–124, 2017.[28] C. Yang, T. Nguyen, and E. Blasch, “Mobile posi-tioning via fusion of mixed signals of opportunity,”
IEEE Aerospace and Electronic Systems Magazine ,vol. 29, no. 4, pp. 34–46, April 2014.[29] Z. Kassas, J. Morales, K. Shamaei, and J. Khalife,“LTE steers UAV,”
GPS World Magazine , vol. 28,no. 4, pp. 18–25, April 2017.[30] M. Maaref, J. Khalife, and Z. Kassas, “Lane-levellocalization and mapping in GNSS-challenged en-vironments by fusing lidar data and cellular pseudo-ranges,”
IEEE Transactions on Intelligent Vehicles ,vol. 4, no. 1, pp. 73–89, March 2019.[31] Z. Kassas, M. Maaref, J. Morales, J. Khalife,and K. Shamaei, “Robust vehicular localizationand map matching in urban environments throughIMU, GNSS, and cellular signals,”
IEEE IntelligentTransportation Systems Magazine , vol. 12, no. 3,pp. 36–52, June 2020.[32] C. Yang and A. Soloviev, “Mobile positioning withsignals of opportunity in urban and urban canyonenvironments,” in
IEEE/ION Position, Location,and Navigation Symposium , 2020, pp. 1043–1059.[33] J. Khalife and Z. Kassas, “Precise UAV navigationwith cellular carrier phase measurements,” in
Pro-ceedings of IEEE/ION Position, Location, and Nav-igation Symposium , April 2018, pp. 978–989.[34] K. Shamaei and Z. Kassas, “Sub-meter accurateUAV navigation and cycle slip detection with LTEcarrier phase,” in
Proceedings of ION GNSS Con-ference , September 2019, pp. 2469–2479.[35] M. Driusso, C. Marshall, M. Sabathy, F. Knutti,H. Mathis, and F. Babich, “Indoor positioning us-ing LTE signals,” in
Proceedings of InternationalConference on Indoor Positioning and Indoor Nav-igation , October 2016, pp. 1–8.[36] C. Gentner, “Channel-SLAM: Multipath assistedpositioning,” Ph.D. dissertation, Ulm University,2018.[37] A. Abdallah, K. Shamaei, and Z. Kassas, “In-door positioning based on LTE carrier phase mea-surements and an inertial measurement unit,” in
Proceedings of ION GNSS Conference , September2018, pp. 3374–3384.[38] A. Abdallah and Z. Kassas, “Evaluation of feedbackand feedforward coupling of synthetic aperture nav-igation with LTE signals,” in
Proceedings of IEEEVehicular Technology Conference , September 2019,pp. 1–6.[39] A. Abdallah and Z. Kassas, “Deep learning-aidedspatial discrimination for multipath mitigation,” in
Proceedings of IEEE/ION Position, Location, andNavigation Symposium , April 2020, pp. 1324–1335.[40] A. Paul and E. Wan, “RSSI-based indoor local-ization and tracking using sigma-point Kalmansmoothers,”
IEEE Journal of Selected Topics in Sig-nal Processing , vol. 3, no. 5, pp. 860–873, 2009.[41] H. Ahn and W. Yu, “Environmental-adaptive RSSI-based indoor localization,”
IEEE Transactions onAutomation Science and Engineering , vol. 6, no. 4,pp. 626–633, 2009.[42] F. Afroz, R. Subramanian, R. Heidary, K. San-drasegaran, and S. Ahmed, “SINR, RSRP, RSSI andRSRQ measurements in long term evolution net-works,”
International Journal of Wireless and Mo-bile Networks , vol. 7, no. 4, pp. 113–123, 2015.[43] X. Ye, X. Yin, X. Cai, A. Yuste, and H. Xu,“Neural-network-assisted UE localization usingradio-channel fingerprints in LTE networks,”
IEEEAccess , vol. 5, pp. 12 071–12 087, 2017.[44] G. Pecoraro, S. Domenico, E. Cianca, and M. Sanc-tis, “CSI-based fingerprinting for indoor localiza-tion using LTE signals,”
EURASIP Journal on Ad-vances in Signal Processing , vol. 2018, no. 1, p. 49,2018.[45] H. Zhang, Z. Zhang, S. Zhang, S. Xu, and S. Cao,“Fingerprint-based localization using commercialLTE signals: A field-trial study,” in
Proceedings ofIEEE Vehicular Technology Conference , 2018, pp.1–5.[46] Z. Yang, Z. Zhou, and Y. Liu, “From RSSI to CSI:Indoor localization via channel response,”
Interna-tional Journal of Wireless and Mobile Networks ,vol. 46, no. 2, pp. 1–32, ACM Computing Surveys.[47] A. Abdallah, K. Shamaei, and Z. Kassas, “Per-formance characterization of an indoor localizationsystem with LTE code and carrier phase measure-ments and an IMU,” in
Proceedings of InternationalConference on Indoor Positioning and Indoor Nav-igation , September 2019, pp. 1–8.[48] K. Shamaei and Z. Kassas, “LTE receiver designand multipath analysis for navigation in urban en-vironments,”
NAVIGATION, Journal of the Instituteof Navigation , vol. 65, no. 4, pp. 655–675, Decem-ber 2018.[49] K. You, “How does learning rate decay helpmodern neural networks?,” arXiv preprintarXiv:1908.01878 , 2019.[50] N. Srivastava, “Dropout: A simple way to preventneural networks from overfitting,”