Deep Learning-based Symbolic Indoor Positioning using the Serving eNodeB
DDeep Learning-based Symbolic Indoor Positioningusing the Serving eNodeB
Fahad Alhomayani and Mohammad Mahoor
Department of Electrical and Computer Engineering, University of Denver, USA [email protected] and [email protected]
Abstract —This paper presents a novel indoor positioningmethod designed for residential apartments. The proposedmethod makes use of cellular signals emitting from a servingeNodeB which eliminates the need for specialized positioninginfrastructure. Additionally, it utilizes Denoising Autoencoders tomitigate the effects of cellular signal loss. We evaluated the pro-posed method using real-world data collected from two differentsmartphones inside a representative apartment of eight symbolicspaces. Experimental results verify that the proposed methodoutperforms conventional symbolic indoor positioning techniquesin various performance metrics. To promote reproducibility andfoster new research efforts, we made all the data and codesassociated with this work publicly available.
Index Terms —Cellular Networks, Dataset, Deep Learning,Denoising Autoencoder, eNodeB, Symbolic Indoor Positioning.
I. I
NTRODUCTION
Interest in indoor positioning research has grown substan-tially in recent years. This can be attributed to the multitudeof potential applications enabled by indoor positioning [1]–[4]. Yet, designing an indoor positioning system has remaineda challenging task since indoor environments are very com-plex and are often characterized by non-line-of-sight settings,moving people and furniture, walls of different densities, andthe presence of different indoor appliances that alter indoorsignal propagation. Among the techniques used for indoorpositioning, location fingerprinting, or simply fingerprinting,has received the most attention due to its simplicity andability to produce accurate positioning estimates [5]. Theconcept of fingerprinting is to identify indoor spatial locationsbased on location-dependent measurable features (i.e., locationfingerprints). Examples of location fingerprints include radiofrequency fingerprints (WiFi [6], Bluetooth [7], cellular [8]),magnetic field fingerprints [9], image fingerprints [10], andhybrid fingerprints [11]. However, the main drawback of fin-gerprinting is the laborious and time-consuming site surveyingtask in which fingerprints are collected at predefined referencepoints (RPs) with known coordinates. Depending on the areato be covered by the system and the accuracy requirement, thenumber of required RPs can be significant. Symbolic position-ing tries to relax this requirement by collecting fingerprintsin zones rather than at points [11]. However, the concept ofdistance is lost since zones are independent and the userslocation is now expressed symbolically (e.g., in the kitchen)instead of physically (using a coordinate system).
Stochastic Corruption
Smartphone eNodeB
TRAININGTESTING
Cellular Data
FingerprintDAE … DAE DAE N DatabaseCompute Reconstruction ErrorsSoftmax FunctionEstimated Symbolic Space
Fig. 1. The general scheme of the proposed method representing the trainingand testing phases.
In this paper, we treat the indoor positioning problem asa classification problem. Each symbolic space in the environ-ment has different cellular signal propagation characteristicsand, hence, should be considered as a unique class. Todistinguish one class from another (i.e., one symbolic spacefrom another), we leverage Denoising Autoencoders (DAEs).The motivation behind employing DAEs, as opposed to otherlearning algorithms, is their ability to handle noisy dataeffectively and efficiently. Our experimental results, whichare based on real signal measurements collected inside aresidential apartment, verify that the proposed method out-performs conventional symbolic indoor positioning techniqueson various performance metrics. The general scheme of theproposed method is depicted in Fig. 1.The remainder of this paper is organized as follows. SectionII reviews some of the recent work in deep learning-basedindoor positioning. Section III describes and validates thedataset used in this study. Section IV provides backgroundon Autoencoders and discusses the design of the proposedmethod. Section V reports on the evaluation experiments andanalyzes the results. Section VI concludes the paper.II. R
ELATED W ORK
In this section, a review of some recent research effortsthat utilize machine learning for symbolic indoor positioningis provided, followed by a review of some recent research a r X i v : . [ ee ss . SP ] S e p fforts that utilize machine learning for cellular-based indoorpositioning. A. Machine Learning for Symbolic Indoor Positioning
Werner et al . [10] utilized the Convolutional Neural Net-work (CNN)-based AlexNet as a generic feature extractor toclassify a query image to one of rooms. No fine-tuningwas performed on the pre-trained network; instead, the authorsdirectly fed the features extracted by the first Fully-Connected(FC) layer to a Nave Bayes classifier. These features helpedtheir model to generalize from local to global views (i.e.,from small views in training to large views in testing) well.However, this did not hold when attempting to generalize fromglobal to local views. Nowicki and Wietrzykowski [6] usedan Autoencoder (AE) followed by an FC network for multi-building and multi-floor classification using WiFi fingerprints.The purpose of the AE is to perform dimensionality reduction.This is important because a WiFi fingerprint has entries forall Access Points (APs) detected in an entire environment, butonly a subset of these APs is observed for different locations.This is especially true for large-scale environments. Mostrecently, Tamas and Toth [11] performed a performance anal-ysis of five machine learning classifiers for symbolic indoorpositioning. They used hybrid fingerprints (WiFi, Bluetooth,and magnetometer) to evaluate and compare the classifiersbeing studied. The fingerprints were obtained from the MiskolcIIS dataset which contains measurements from zones ofdifferent sizes inside a three-story university building.Our proposed method has several advantages compared tothe aforementioned works: • It preserves privacy because it does not require thecapturing of images for positioning. • It is well-suited for small-scale indoor environments,namely residential apartments and homes, where peoplespend most of their time. • It does not require on-premises infrastructures such asWiFi APs or Bluetooth beacons for operation. Instead, itrelies on omnipresent cellular signals. • It has little overhead because only a single fingerprinttype is required for positioning which eliminates thecomplexity associated with fusing multiple fingerprinttypes.
B. Machine Learning for Cellular-based Indoor Positioning
Rizk et al . [8] used an FC network to perform cellularReceived Signal Strength (RSS) fingerprinting. Data augmen-tation techniques were used to increase the training set by -fold. The authors achieved a positioning error of less than of the time. However, to achieve this accuracy, the RSSfrom Second-Generation ( G) cellular Base Stations (BSs)had to be measured. Later, the authors used a Recurrent NeuralNetwork (RNN) to capture the temporal dependency betweenconsecutive RSS measurements received from the serving BS[12]. The achieved positioning accuracy was comparable tothat acquired by their previous approach, however, a mea-surement window of at least three seconds had to be fed to the RNN. Arnold et al . [13] used a custom-built linear arrayof Multiple-Input Multiple-Output (MIMO) antennas installedin a
20 m by area for indoor positioning. They used anFC network to correlate the antennas’ channel coefficients toa D position relative to the array’s location. Vieira et al .[14] used a CNN to learn the structure of massive MIMOchannels for indoor positioning. A cellular channel modelwas used to generate unique channel fingerprints for eachtraining/testing position. These fingerprints represent clustersof multipath components obtained from a BS equipped with alinear array of antennas. Since all measurements were basedon simulated data, real-world measuring impairments such asnoise and channel fading were not considered.Compared to the previous works, our proposed methodcombines several features that place it in a unique position: • It employs DAEs to handle incomplete measurementscaused by unpredictable cellular signal loss. • It only utilizes the information measured from the servingBS, which is a Fourth-Generation ( G) cellular BS (alsoknown as an eNodeB in Long-Term Evolution). • It is well-suited for real-time positioning applications,given the parametric nature of DAEs, in addition toperforming one-shot positioning (i.e., only a single fin-gerprint sample is required to estimate the users location). • It is based on real-world measurements emitting from areal eNodeB. No simulated, interpolated, or augmenteddata were used in this study. • It is well-suited for smartphone-based positioning be-cause all measurements were collected using smartphonesas opposed to custom-built collection platforms.III. D
ATASET D ESCRIPTION
Nearly all indoor positioning solutions found in the lit-erature were evaluated using private data. Thus, the resultsobtained are self-reported and cannot be reproduced. Addi-tionally, the lack of publicly available datasets that can beused to develop, evaluate, and compare indoor positioningsolutions constitutes a high entry barrier for studies. For thesereasons, we made the dataset used in this study publiclyavailable [15]. The following subsections describe the datacollection platform, environment, procedure, and technicalquality, respectively.
A. Data Collection Platform
We used two smartphones for data acquisition: SamsungsGalaxy S + (Phone ) and Googles Pixel (Phone ).Both smartphones ran on Android . The motivation be-hind choosing Android-powered smartphones was twofold.First, Android provides Application Programming Interfaces(APIs) that allow for acquiring raw data at the hardwarelevel. Second, Android-powered smartphones account for over
74 % of the market share worldwide [16]. We attached thetwo smartphones to a tripod using a dual mount (Fig. 2).Both smartphones were in portrait mode and were kept ata fixed height of
130 cm . The tripod head was adjusted totilt the smartphones at a ∼ -degree angle to the vertical ig. 2. A picture of Phone 1 and Phone 2 attached to the tripod plane. We installed the same third-party app [17] used forthe data collection on both smartphones. The app allowed forconveniently collecting and exporting cellular network data. B. Data Collection Environment
We performed data collection inside a residential apart-ment of eight symbolic spaces. As seen from the apart-ments layout (Fig. 3), the symbolic spaces include a liv-ing room ( . × . ), a sunroom ( . × . ), a bed-room ( . × . ), a hallway ( . × . ), a diningroom ( . × . ), a kitchen ( . × . ), a bathroom( . × . ), and a walk-in closet ( . × . ). The floorplan delineating the apartments dimensions is provided along-side the dataset in the form of a .vsdx file. C. Data Collection Procedure
A smartphones cellular modem constantly scans the cellularnetwork for cell selection/reselection and handover purposes.Android provides APIs to extract data associated with scanssuch as Reference Signal Received Power (RSRP) and cellidentity information [18]. For each of the symbolic spacesdescribed above, we collected minutes of cellular data (perphone) at a sampling rate of hertz ( Hz ). During data collec-tion, we systematically changed the position and orientationof the tripod to uniformly cover space and direction. Samplingresults were exported as a .csv file and named with the smart-phones and spaces name (e.g., Phone2_Bedroom.csv ).Table I lists all fields in each data sample and their descrip-tions. As an example, Fig. 4 plots the data collected from thesmartphones located inside the walk-in closet.
Living room SunroomHallway Bedroom Walk-in closet Dining room Kitchen Bathroom storage storage storageW
D boiler room
Fig. 3. Layout of the apartment where data was collected d B m outlierPhone 1 Phone 2 d B data loss Time ( s ) Time ( s ) Fig. 4. Plots of cellular data showing examples of outliers and data loss inthe data collected by Phone and Phone inside the walk-in closet. D. Technical Validation
The technical quality of the dataset was evaluated usingexperiments that consider its reliability and validity:
1) Measurement Reliability:
Before the collection cam-paign, we captured cellular data over three different days at thesame location. We used Spearmans and Kendalls correlationcoefficients to quantify the degree of consistency betweentemporal measurements for a given phone. Table II showsSpearman’s and Kendall’s correlation coefficients for the twosmartphones for all possible pairs of days. Given that correla-tion results are high (i.e., close to the maximum value of . ),it can be concluded that the dataset possesses a high degreeof reliability.
2) Measurement Validity:
We assessed measurement valid-ity by comparing the cellular data captured by the two phonesand checking for consistency. Accordingly, for a given day,we used Spearman’s and Kendall’s correlation coefficientsto quantify the degree of consistency between the measure-ments obtained by the phones. The correlation results for theforegoing three days are shown in Table III. These resultsdemonstrate high levels of consistency, which attests to thevalidity of the dataset.IV. B
ACKGROUND AND P ROPOSED M ETHOD
A. Autoencoders
Autoencoders (AEs) are a family of feedforward neuralnetworks that have been used in unsupervised learning tasks.AEs have the same number of neurons in the input layer asthe output layer. A typical AE is trained to reconstruct aninput without memorizing or directly copying it. Instead, anencoder-decoder approach is used, as seen in Fig. 5. Thishourglass-shaped architecture forces the network to encode(compress) the input into a latent code from which the inputcan be decoded (reconstructed). Backpropagation is used tolearn the network’s parameters by minimizing reconstructionloss between the input and the reconstructed input. Commonvariants of AEs are Denoising AEs (DAEs). DAEs are trainedto reconstruct an input from a corrupted version of it (Fig. 5).
ABLE IF
IELD LABELS OF DATA SAMPLES AND THEIR DESCRIPTION Date_Time
The date and time the sample was captured as
YYYYMMDDhhmmss PLMN_ID
The Public Land Mobile Network (PLMN) IDentifier3 eNodeB_ID
The E-UTRAN NodeB (eNodeB) IDentifier that is used to uniquely identify an eNodeB within a PLMN4
Cell_ID
The Cell IDentifier which is an internal descriptor for a cell. It can take any value between and ECI
The E-UTRAN Cell Identifier that is used to uniquely identify a cell within a PLMN.
ECI = × eNodeB_ID + Cell_ID .6 RSRP
The Reference Signal Received Power in decibel-milliwatts ( dBm )7 RSRQ
The Reference Signal Received Quality in decibels ( dB )8 SINR
The Signal to Interference and Noise Ratio in dB UMTS_neighbors
The number of neighboring Universal Mobile Telecommunications Service (UMTS) cells10
LTE_neighbors
The number of neighboring Long-Term Evolution (LTE) cells11
RSRP_strongest
The Reference Signal Received Power, in dBm , corresponding to the strongest neighboring LTE cell
TABLE IIR
ESULTS OF THE CORRELATION ANALYSIS BETWEEN THE MEASUREMENTS OBTAINED ON THREE DIFFERENT DAYS FOR P HONE AND P HONE . T HERESULTS WERE GENERATED USING SYNCHRONIZED READINGS OF FIELDS – . T HE p - VALUES OF ALL RESULTS WERE LESS THAN . . Phone 1 Phone 2 { Day , Day } { Day , Day } { Day , Day } { Day , Day } { Day , Day } { Day , Day } Spearman’s ρ .
992 0 .
988 0 .
981 0 .
990 0 .
976 0 . Kendall’s τ .
983 0 .
973 0 .
956 0 .
977 0 .
945 0 . TABLE IIIR
ESULTS OF THE CORRELATION ANALYSIS BETWEEN THEMEASUREMENTS OBTAINED FROM P HONE & P HONE FOR THREEDIFFERENT DAYS . T
HE RESULTS WERE GENERATED USING READINGS OFFIELDS – . T HE p - VALUES OF ALL RESULTS WERE LESS THAN . . Phone 1 & Phone 2
Day Day Day Spearman’s ρ .
991 0 .
995 0 . Kendall’s τ .
979 0 .
989 0 . input 𝑥 output 𝑥̃ 𝑥 ≈ 𝑥̃ encoder decoder bottleneck layer latent code corrupted input Autoencoder
Denoising Autoencoder
Fig. 5. The structure of an AE and a DAE
B. Proposed Method
The design of the proposed method is inspired by thesuccessful application of AEs for anomaly detection [19]:An AE, when solely trained on normal data instances, failsto reconstruct abnormal data instances, hence, producing alarge reconstruction error. The data instances that produce highresidual errors are considered outliers.The proposed method takes a normalized cellular datasample captured from the serving eNodeB as input (1) andproduces an output (2) that is a probability distribution overthe set of symbolic spaces in the environment: X = (cid:2) RSRP , RSRQ , SINR , UMTS_neighbors , LTE_neighbors , RSRP_strongest (cid:3) ; X ∈ R : { x i ∈ R | ≤ x i ≤ } (1) Y = (cid:2) Pr( space | X ) , · · · , Pr( space n | X ) (cid:3) ; Y ∈ R n : { y i ∈ R | y + · · · + y n = 1 } (2) Given DAEs’ data-driven learning ability, the proposedmethod does not make any assumptions about feature inde-pendence or the nature of the boundary separating the classes.The input vector is corrupted to emulate a randomizedloss of cellular data. This is accomplished using a Hadamardproduct of (1) and an all-ones vector ( (cid:126) ) whose elements arerandomly set to with a predefined probability p loss . Forexample, if p loss is set to . , there is a
50 % chance thata given field entry will be set to zero.During the training phase, a dedicated DAE is employedfor each symbolic space. Each DAE is solely trained on thedata collected at its corresponding symbolic space. By follow-ing this training strategy, we expect that, during the testingphase, all DAEs, except for one, will generate a relativelyhigh reconstruction error when fed the same testing sample.Consequently, the symbolic space associated with the DAEthat generated the lowest reconstruction error is consideredas the estimated symbolic space. To construct (2), we used aSoftmax function (3) during the testing phase:
Pr( space i | X ) = S ( L i ) = exp(1 / L i ) (cid:80) ni =1 exp(1 / L i ) (3) where L i is the reconstruction loss generated by the i th DAE.
1) DAE Architecture:
All DAEs have the same architecturewhich consists of an input layer of neurons, a hidden layer(and its mirror layer) of neurons, a bottleneck layer of neurons, and an output layer of neurons. We developedthe DAEs using Keras with the hyperparameters listed in TableIV. We selected these hyperparameters using grid search andcross-validation. We used early stopping and dropout to avoidoverfitting. ABLE IVH
YPERPARAMETER TUNING
Hyperparameter Value
Batch size
Dropout rate . Optimizer Adadelta ( ρ = 0 . , (cid:15) = 1e − )Learning rate . Activation function ReLUEpochs , Loss function Binary cross-entropyWeight initializer Xavier uniformBias initializer Zeros
2) Dataset Preprocessing:
From each symbolic space, therewere , samples collected by each smartphone. For theentire collection period, and throughout the collection envi-ronment, Phone and Phone camped on the same LTEcell (i.e., ECI:98059528 ). Thus, entries for field labels – were identical for all samples. For a given symbolicspace, we combined the samples collected by Phone andPhone to create a single dataset for training and testing thecorresponding DAE. After the samples in the combined datasetwere randomly shuffled, we allocated
80 % of them for trainingand validation, and the remaining
20 % for testing. Since inputfeatures are measured in different units, their values werenormalized between and . This was performed after thedataset was split to avoid data contamination. Fig. 1 showsthe general scheme of the proposed method.V. E XPERIMENTS AND R ESULTS
This section evaluates the performance of the proposedmethod and investigates the impact of p loss and device hetero-geneity on positioning accuracy. Associated computing scriptsare publicly available in our figshare repository [15]. A. Performance Evaluation
We trained the proposed method with a p loss value of . applied to the training set. The metrics used for performanceevaluation are Accuracy, Precision, Recall, and F -score asdefined in [20]. We compared the performance of the proposedmethod against classifiers that are extensively used for indoorpositioning, namely k -Nearest Neighbor ( k -NN) and SupportVector Machine (SVM). For the sake of fair comparison, wetrained these classifiers on the same training set used forthe proposed method and fine-tuned their parameters usinggrid search and cross-validation. The testing set used forcomparison was contaminated with a p loss value of . .Fig. 6 reports on the classification results and shows theconfusion matrices of the three methods. The results clearlyshow that the proposed method outperforms both k -NN andSVM on all metrics. As mentioned earlier, both smartphonesconnected to the same LTE cell throughout the environment.However, it is possible, depending on network parameters, thata connection alternates between multiple cells. Incorporatingthe information obtained by additional cells is expected tofurther enhance performance because location discernibilitywill increase with increased features. Interesting observations can be made by examining the con-fusion matrices in Fig. 6. For example, higher degrees of con-fusion tend to occur between symbolic spaces that are close toeach other (e.g., Kitchen and Bathroom or Sunroom and Livingroom). Nevertheless, observations exist that prove contrary tothis assumption. For instance, there is low confusion betweenBedroom and Walk-in closet despite their proximity. In fact,it is more likely to confuse Bedroom for Dining room than itis to confuse Bedroom for Walk-in closet. Such observationscould be the result of the complex changes that cellular signalsundergo when propagating indoors. Confirming this conjectureis a topic of future research. B. Effect of p loss on Accuracy To study the impact of p loss on positioning accuracy, weevaluated the proposed method, k -NN, and SVM using datacontaminated with varying p loss values. More specifically, wegenerated copies of the testing set and contaminated eachcopy with a different p loss value that ranged from . to . ,using . increments. The methods’ accuracy scores thatcorresponded to each p loss value were recorded and plottedin Fig. 7. One observation that can be made from Fig. 7 isthat, as loss increases, so does the performance gap betweenthe proposed method and k -NN/SVM. This primarily suggeststhat DAEs learned more robust features than k -NN and SVM. C. Effect of Device Heterogeneity on Accuracy
Smartphones obtain cellular data readings from their cellularchipsets. Since these chipsets are manufactured by differentvendors, hardware and firmware specifications are not uniformacross smartphones. This results in heterogeneous receptioncharacteristics which, in turn, can degrade the accuracy of thepositioning system [21].In this experiment, we investigated device heterogeneityby training the proposed method, k -NN, and SVM on thedata obtained from one smartphone and testing on the dataobtained from 1) the same smartphone and 2) the othersmartphone to quantify the difference in performance. TableV reports on the experiments results. As clearly seen fromTable V, device heterogeneity is a significant problem in allthree methods. Average accuracy drops of . , . , and . are observed in the proposed method, k -NN, and SVM,respectively. In the field of indoor positioning, there is ongoingresearch regarding overcoming device heterogeneity and weintend to address this limitation in subsequent research. TABLE VR
ESULTS OF THE DEVICE HETEROGENEITY ANALYSIS . p loss IS SET TO . FOR BOTH TRAINING AND TESTING . Training Data Testing Data (accuracy)
Phone 1 Phone 2 Phone 1 Phone 2Proposed Method (cid:88) .
565 0 . (cid:88) .
299 0 . k -NN (cid:88) .
395 0 . (cid:88) .
230 0 . SVM (cid:88) .
396 0 . (cid:88) .
222 0 . a t h r oo m B e d r oo m D i n i n g _ r oo m H a ll w a y K i t c h e n L i v i n g _ r oo m S un r oo m W a l k - i n _ c l o s e t BathroomBedroomDining_roomHallwayKitchenLiving_roomSunroomWalk-in_closet T r u e S y m b o li c S p a c e Proposed Method (Accuracy=0.477 Precision=0.475Recall=0.477 F1-score=0.474) B a t h r oo m B e d r oo m D i n i n g _ r oo m H a ll w a y K i t c h e n L i v i n g _ r oo m S un r oo m W a l k - i n _ c l o s e t Predicted Symbolic Space
BathroomBedroomDining_roomHallwayKitchenLiving_roomSunroomWalk-in_closet k -NN (Accuracy=0.355 Precision=0.355Recall=0.355 F1-score=0.354) B a t h r oo m B e d r oo m D i n i n g _ r oo m H a ll w a y K i t c h e n L i v i n g _ r oo m S un r oo m W a l k - i n _ c l o s e t BathroomBedroomDining_roomHallwayKitchenLiving_roomSunroomWalk-in_closet
SVM (Accuracy=0.361 Precision=0.362Recall=0.361 F1-score=0.361)
Fig. 6. Accuracy, Precision, Recall, F -score, and the confusion matrix of symbolic space prediction of the proposed method, k -NN, and SVM. p loss a cc u r a c y s c o r e Proposed Method k -NNSVM Fig. 7. The effect of p loss on accuracy for the proposed method, k -NN, andSVM. VI. C
ONCLUSION
This paper presented the design and evaluation of a novelcellular-based symbolic indoor positioning method. At itscore, the proposed method utilizes DAEs to alleviate the ef-fects randomized signal loss has on positioning. Experimentalresults demonstrated that the proposed method outperformsconventional methods with respect to several performancemetrics. In its current form, the proposed method does notconsider temporal information/patterns in location inference.Exploiting temporal dependencies among data samples helpsmaintain temporal coherence which could further enhanceperformance. In this respect, we intend to extend this work byintegrating deep learning architectures suitable for processingtime-series data (e.g., RNN Autoencoders [22]).R
EFERENCES[1] D. Macagnano et al. , “Indoor positioning: A key enabling technologyfor iot applications,” in , 2014, pp. 117–118.[2] P. Rashidi and A. Mihailidis, “A survey on ambient-assisted living toolsfor older adults,”
IEEE J. Biomed. Health Inform. , vol. 17, no. 3, pp.579–590, 2013.[3] M. Werner,
Indoor location-based services: prerequisites and founda-tions . Springer, 2014.[4] A. F. G. Ferreira et al. , “Localization and positioning systems foremergency responders: A survey,”
IEEE Commun. Surveys Tuts. , vol. 19,no. 4, pp. 2836–2870, 2017.[5] F. Alhomayani and M. H. Mahoor, “Deep learning methods forfingerprint-based indoor positioning: a review,”
Journal of LocationBased Services , vol. 0, no. 0, pp. 1–72, 2020. [6] M. Nowicki and J. Wietrzykowski, “Low-effort place recognition withwifi fingerprints using deep learning,” in
Int. Conf. Automat.
Springer,2017, pp. 575–584.[7] R. Faragher and R. Harle, “Location fingerprinting with bluetooth lowenergy beacons,”
IEEE J. Sel. Areas Commun. , vol. 33, no. 11, pp.2418–2428, Nov 2015.[8] H. Rizk et al. , “Cellindeep: Robust and accurate cellular-based indoorlocalization via deep learning,”
IEEE Sensors J. , vol. 19, no. 6, pp.2305–2312, March 2019.[9] F. Al-homayani and M. Mahoor, “Improved indoor geomagnetic fieldfingerprinting for smartwatch localization using deep learning,” in , Sep. 2018,pp. 1–8.[10] M. Werner et al. , “Deepmovips: Visual indoor positioning using transferlearning,” in , 2016, pp. 1–7.[11] J. Tamas and Z. Toth, “Classification-based symbolic indoor positioningover the miskolc iis data-set,”
Journal of Location Based Services ,vol. 12, no. 1, pp. 2–18, 2018.[12] H. Rizk and M. Youssef, “Monodcell: A ubiquitous and low-overheaddeep learning-based indoor localization with limited cellular informa-tion,” in
Proc. 27th ACM SIGSPATIAL Int. Conf. Advances in Geo-graphic Inf. Syst. , 2019, pp. 109–118.[13] M. Arnold et al. , “On deep learning-based massive mimo indoor userlocalization,” in , 2018, pp. 1–5.[14] J. Vieira et al. , “Deep convolutional neural networks for massive mimofingerprint-based positioning,” in , Oct 2017, pp. 1–6.[15] F. Alhomayani and M. H. Mahoor, “Cellular networkmeasurements for symbolic indoor positioning,” Sep 2020. [On-line]. Available: https://figshare.com/articles/dataset/Cellular NetworkMeasurements for Symbolic Indoor Positioning/13010387/1[16] “Mobile operating system market share worldwide (dec 2018 -dec 2019),” statCounter. [Online]. Available: https://gs.statcounter.com/os-market-share/mobile/worldwide[17] “Netmonitor pro,” vitaly V. [Online]. Available: https://play.google.com/store/apps/details?id=ru.v a v.netmonitorpro&hl=en US[18] “android.telephony,” android Developers. [Online]. Available: https://developer.android.com/reference/android/telephony/package-summary[19] R. Chalapathy and S. Chawla, “Deep learning for anomaly detection: Asurvey,” arXiv preprint arXiv:1901.03407 , 2019.[20] “Classification metrics.” [Online]. Available: https://scikit-learn.org/stable/modules/model evaluation.html et al. , “Modern wlan fingerprinting indoor position-ing methods and deployment challenges,”
IEEE Commun. Surveys Tuts. ,vol. 19, no. 3, pp. 1974–2002, thirdquarter 2017.[22] N. Srivastava et al. , “Unsupervised learning of video representationsusing lstms,” in