Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

1 Article

2 Smart adaptive positioning based on multiple


3 wireless technologies in indoor scenarios
4 Jordi Mongay Batalla 1,*, Jose Angel Flores 1, Constandinos X. Mavromoustakis 2 and George
5 Mastorakis 3
6 1 National Institute of Telecommunications and Warsaw University of Technology, Poland;
7 j.mongay@itl.waw.pl
8 2 University of Nicosia, Cyprus; mavromoustakis.c@unic.ac.cy

9 3 Technological Educational Institute of Crete, Greece; gmastorakis@ieee.org

10 * Correspondence: jordim@tele.pw.edu.pl Tel.: +48-22-234-7886

11 Received: 21 September 2018; Accepted: date; Published: date

12 Abstract: General stores aim to follow clients’ activity, in order to increase their satisfaction, as well
13 as to adopt optimized decisions for the management of the shop and for marketing purposes.
14 Current proposals collide with variability of environmental conditions, so that the solutions worked
15 in stationary systems but they had real difficulties to be efficient in real environments with high
16 variability. This paper fills in the existing gap between static approaches and dynamic indoor
17 positioning systems, by presenting a solution for adapting the system to changing conditions of the
18 environment. The proposed system assumes fingerprint calibration in presence of clients, which
19 introduce variations in the electromagnetic environment, so that the positioning algorithm takes
20 into account such variations. This paper also covers a full solution for indoor positioning system
21 that reduces complexity of hardware, so that a multi-standard-transceiver infrastructure may be
22 adopted with reduced capital and operational expenditures. We developed the system from the
23 scratch and performed extensive testbed experiments, which show that the multi-technology
24 transceiver feature allows to increase positioning accuracy, as well as to include permanent
25 fingerprints calibration.

26 Keywords: Positioning system; Bluetooth; Received Signal Strength; Mahalanobis distance;


27 calibration.
28

29 1. Introduction

30 1.1. Motivation
31 General stores are following the steps of the online market, which collects statistics from Web
32 customers and their profiles mainly for advertising purposes. Shopping business has reached similar
33 needs not only for marketing purposes but also for increasing client’s satisfaction, as well as for
34 adopting optimized decisions for management. In addition, general stores require regular and exact
35 monitoring of customers’ behavior, to allow the manager take proper and effective decisions in the
36 administration of resources. The interesting information for shop managers is, among others, the
37 quantity of customers in the shop (variable in time), the interests of the customers in the shop
38 (elements or products that may catch the customers’ attention) and customer profile (relation
39 between interests and real purchase). Such information allows an effective distribution of the space
40 in the shop and placement of products. For example, the products offered in the most visited
41 corridors in the shop could be charged with a share for placement attractiveness. Currently, shops
42 rely on one of the following solutions for collecting consumers’ behavior data:

Sensors 2018, 18, x; doi: FOR PEER REVIEW www.mdpi.com/journal/sensors


Sensors 2018, 18, x FOR PEER REVIEW 2 of 23

43 • Customer monitoring provided by shop staff. This is a very questionable solution due to
44 the fact that the staff has many other duties than data collection, which in practice implies that this
45 task has very low priority. As a result, the number of errors in the methodology and monitoring
46 process is high and the results are not reliable.
47 • Monitoring provided by external companies based on interviews to the customers. External
48 experts perform data collection of customers’ behavior. This solution offers better results. However,
49 it is full of inaccuracy and inconsistences due to the quiz-character of the methodology. Moreover,
50 this solution increases costs due to subcontracts.
51 • External reports on customers’ behavior. These reports are, in general, unreliable, since they
52 are based on general analysis not related with given shop and situation.
53 Since the current solutions do not provide enough data reliability (both tools and methodologies
54 result extremely inaccurate), other approaches relying on technological solutions have been proposed
55 in the last years. However, such solutions failed to be implemented in real shops since the current
56 level of development makes them valid only for static environmental conditions.
57 This paper presents an advanced system for collecting and analyzing data on the behavior of
58 people in a given space, which may reach expectations of potential customers and may increase their
59 degree of satisfaction. Customers’ activity will be obtained in the analyzed space via the information
60 received through the wireless access networks, which receive radio signal from the devices owned
61 by the customers. The novelty of this solution compared to other research studies is that the proposed
62 system is capable of adapting to change conditions of the environment, so that it may overcome the
63 difficulties for implementing the solution in general stores. Adaptation to changing environmental
64 conditions is achieved thanks to the on-line calibration of the system, which allows to adjust the
65 system setting-up parameters to the current state of the analyzed space (considering new objects in
66 the space causing interferences, echoes, signal shadows, etc.). The presented Business Intelligence
67 solution relies on fingerprints of Received Signal Strength (RSS) measurements of the signal sent by
68 the customers’ devices and received by multiple transceivers located within the analyzed space.
69 Fingerprints map is updated in time, such that the map includes information about the current
70 situation of the analyzed space and takes into consideration the effects of new obstacles and
71 interferences of the signal.
72 Another innovation is achieved to the integration of relatively low-cost multi-transceiver multi-
73 technology system for positioning purposes. As a result, the presented system reaches localization
74 accuracy at the range of 0.4 meters, while maintaining limited costs (Capital Expenditures) and
75 effective power consumption (low Operational Expenditures). Concretely, we propose a new
76 approach to multi-transceiver system that consists of the separation of radio signal processing from
77 message processing. The first one is almost technology-independent (considering the technologies
78 working on 2.4 GHz band), whereas the second functionality depends on the technology. This way,
79 our system operates with different technologies (Wi-Fi, Bluetooth, RFID and others based on beacons)
80 and is open to other technologies, which could be introduced only by developing transmission in the
81 appropriate frequencies. Device positioning process is developed into the technology-dependent
82 modules, allowing for the implementation of different technology-dependent fingerprints map.
83 Moreover, this approach allows to introduce sleep-mode in the HW part of the system whenever
84 location is established. This is a big difference with current systems containing fully-functional
85 transceivers, which cannot be deactivated during their lifecycle as they serve connectivity.
86 Finally, our solution assumes not required action in the positioned devices, simultaneously
87 positioning in several radio technologies operating in the 2.4 GHz band, measurements provided in
88 the environment of a dense network of transceivers (which improves accuracy of measurements),
89 adaptation to dynamic changes in the environment thanks to the integration of calibration and
90 positioning phases and, at last but not least, it reaches limited impact of radio interferences (between
91 technologies and between devices).
92 A strength of our research is, undoubtedly, the development of the full system, which may work
93 on operational conditions of a shop. Based on this implementation, we conducted extended
Sensors 2018, 18, x FOR PEER REVIEW 3 of 23

94 experiments, which aimed to demonstrate the system accuracy in operational conditions after step-
95 by-step calibration phase.

96 1.2. System requirements


97 In this paper we present a comprehensive telecommunication system enabling the collection,
98 analysis and reporting of information about the behavior of people in the study, monitored space.
99 The requirements for such a system are the following:
100 • maximizing simplicity of solution,
101 • minimizing the necessary investment on the user side,
102 • lack of necessity of the activity on the side of the monitored entities,
103 • achieve positioning precision of less than 0.5 m,
104 • maximizing the number of devices, which may be localized and positioned by the system.
105 In practice, this means that the system should work with a wide range of different technologies.
106 From the above requirements, we designed a system with the following assumptions:
107 • The activity at the premises would be monitored using available at the premises network –
108 WI-FI, RFID, IEEE 802.15.4, Bluetooth. Moreover, the system will be open to other and new
109 technologies, such that the integration of new technologies will be as easy as possible. It is worth to
110 pay attention to the legitimacy of the use of RFID technology. For most applications, this technology
111 is used to identify the objects marked with RFID tag, for short distances (up to tens of centimeters).
112 This type of solution, because of the low range and a different scope of used frequencies are not
113 suitable from the point of view of our proposed positioning system. In the present context, we see
114 RFID systems operating in the 2.4 GHz band using active tags (i.e. with their own power supply),
115 which can be used e.g. for monitoring employees.
116 • The function of a transceiver transmitting a locational signal will be carried out by mobile
117 devices used by people entering the room, allowed items lent entrants, e.g. IDs or shopping baskets.
118 • The function of a receiving device, collecting and analyzing locating information will be
119 carried out by multi-technology (at the under layer) controllers.
120 • Full positioning of people and determination of their activity will take at least 2 or 3
121 monitoring devices. This refers to the flat single plane of a typical shape within it is possible to present
122 a three-dimensional XYZ space on a two-dimensional projection in the XY plane (no need to monitor
123 activity in the vertical).
124 • The system will be based on the reception of customer’s device signal in an anonymous
125 way; this is, we are not interested in profiling the users, however, the signal should be properly
126 identified for mapping the customers’ activity. We assume that MAC address is unique into the
127 analyzed space and it does not change during the time when a customer remains in the analyzed
128 space. Let us remark that this is a realistic assumption since only new mobile devices change the
129 MAC address automatically in long time intervals. In additional, it will be possible to use other
130 distinguishing elements, for example: signal transmitters in RFID technology installed, e.g. on a
131 basket shop floor (integration of technology allows for elimination of errors or a combination of
132 information, e.g. abandoning the cart, movement in the office without ID).
133 • The primary function will be data collection, compilation and analysis in a computing
134 platform system on the behavior of people entering the premises, which will allow for (1) good
135 statistics keeping and studying the activity of visitors; (2) study of usability and attractive of a
136 premise. For example, drawing activity paths of customers in the store – creating a heatmap of a room
137 showing the density of tracks in the period, including the possibility of analysis of different time
138 intervals; (3) the process of reporting, based on interest rates, e.g. retention and bounce frequency of
139 visits and tracking indicators for the segments; and (4) verifying marketing objectives with the actual
140 data, e.g. examine the reception exposed offer.

141 1.3. Main contributions


142 The majority of solutions relying on technology-based monitoring of users’ devices indoor
143 positioning makes use of the 802.11 network (Wi-Fi) [1-2]. The advantage is its relative low cost
Sensors 2018, 18, x FOR PEER REVIEW 4 of 23

144 compared to other wireless technologies, and further it is important that in many locations Wi-Fi
145 network already exists as part of the infrastructure. In these cases, it is possible to avoid costly and
146 time-consuming process of infrastructure development. Although Wi-Fi standard during the
147 development did not include functional positioning, its radio signals can be used to assess the
148 location based on the received signal strength (RSS), if localized object is "seen" from at least three
149 access points. This positioning system can be relatively easily implemented to any devices that
150 support Wi-Fi (e.g. smartphones).
151 We can distinguish two basic types of positioning systems in rooms, which are using Wi-Fi
152 networks and a power measurement signal. The first one is the type of proximity (proximity
153 detection), which uses the RSS measurement and propagation models of the radio signal [3-5] [40] to
154 identify the exact location. Development of an accurate propagation model, for each of Wi-Fi access
155 point (AP), in the room in a real environment is extremely difficult. Therefore, the most solutions are
156 burdened with relatively high accuracy positioning error [5-6]. Another common type of localization
157 based on the RSS measured is the positioning using a method called fingerprinting [6-11] [41]. It is
158 based largely on the use of empirical data. Under this method, the localization is usually carried out
159 in two phases: the calibration phase and positioning phase. In the calibration phase, the APs measure
160 the RSS of the signal emitted by a testing device. The average of several measurements is commonly
161 used as a reference value of the RSS. Each of the tests becomes the point of the radio map in which
162 individual locations are defined by geographical coordinates and the specific value of feed for each
163 AP. During positioning phase, the mobile device measures the value of RSS in an unknown location
164 and uses an algorithm to estimate location using previously created radio maps. Because the rooms
165 have unique features of the propagation of the signal, it can be assumed that each location can be
166 determined by a unique combination of RSS.
167 This approach provides a fairly accurate positioning, even in very complex environments, where
168 modeling a complex propagation signal is practically unreachable. Importantly, the fingerprinting
169 techniques usually do not require a precise knowledge of access points location.
170 Considering the accuracy of the measurement method of fingerprinting, the key element is the
171 correlation between the result of RSS measurement and individual items on the radio map. In
172 practice, it comes down to determine the distance between the two abovementioned points, which in
173 statistical terms maps contributions of individual components and uses the correlation between them.
174 In this context, it is essential to select the appropriate measure of distance, as pointed out by the
175 authors in [12], because positioning accuracy is closely related to it. The authors of [13] analyze two
176 different distance measurements in terms of their application to positioning fingerprinting using Wi-
177 Fi networks. These were the Euclidean distance and Mahalanobis distance [14-16]. The tests carried
178 out showed that the highest accuracy can be achieved using Mahalanobis distance. The confirmation
179 of this thesis is usage of the Mahalanobis distances by several other solutions [17-20].
180 A number of enhancements designed to increase accuracy were proposed in these solutions [21].
181 These extensions generally concern the inclusion, in the algorithms, of the volatility of the RRS values
182 for the same location, which often occurs in real conditions. It should be added that a number of
183 solutions for positioning using the method of fingerprinting in order to increase the accuracy of
184 measurement employs also use data received from the Inertial Measurement Unit (IMU) module [22],
185 which is an accelerometer or magnetometer and built a device localized. The advantage of the use of
186 this module is to compensate for the inaccuracy of measurement associated with the movement of
187 the user during the measurement procedure. Measurement data indicating RSS signal strength are
188 combined with the data in IMU module and sent to an application, which on the basis of a specific
189 algorithm calculates the final location [23]. However, it is worth mentioning that in this approach the
190 measurement data is obtained on the customer’s device (active participation on the measurements),
191 which consent might exclude the possibility of using this method for informative purposes of the
192 vendor (which is possible by passive measurements).
193 One drawback of fingerprints technique is that calibration and positioning are decoupled, which
194 results in different conditions of the two processes, due to, e.g., multi-device interferences, changes
195 in indoor furniture position, etc. In fact, such changes (environment) have much higher impact in the
Sensors 2018, 18, x FOR PEER REVIEW 5 of 23

196 efficiency of the system than other aspects as, for example, time. The measurements of the signal
197 taken in different moments maintain uniform, however, small changes in the environment cause
198 serious reduction of positioning efficiency [24]. Therefore, in this paper we propose to joint
199 calibration and positioning and dynamically adapt fingerprints map during the functioning of the
200 positioning system.
201 In the case of Bluetooth technology construction of the positioning platform can be based on
202 various measuring metrics which provide input data for use in further positioning algorithms. There
203 are four main measurement metrics: RSS, LQ (Link Quality), TPL (Transmission Power Level) and
204 IRR (Inquiry Response Rate). As in the case of Wi-Fi, the measure of RSS parameter is the most
205 common and it is used in the abovementioned fingerprinting method. Another parameter possible
206 to obtain in Bluetooth technology is LQ connection status. Tests have shown a dependence of LQ
207 parameter on distance between a Bluetooth transmitter and receiver which makes it possible to use
208 the method based on the above parameter in determining the location. It should be noted that LQ
209 parameter is possible to measure at a time when the customer’s device has an active connection with
210 the master device, which is a significant reduction in the application. In Bluetooth technology, it is
211 also possible to measure the power of the TPL signal transmission. This parameter is possible to
212 specify only in a connected state, just like in the case of LQ. However, some tests [25] have shown a
213 weak dependence of this parameter on the distance, which actually eliminates any method of
214 positioning based on TPL. IRR is the last parameter used to positioning in Bluetooth technology and
215 it indicates the number of responses received in the time interval to inquiries generated by the master
216 in the call establishing process. The parameter is mainly used in a method of fingerprinting; however,
217 tests have shown lower accuracy of location measurement based on this parameter than
218 measurements based on RSS. Therefore, also in Bluetooth device positioning, RSS is the most suitable
219 approach. In Bluetooth Low Energy (BLE) technology, the conclusions are similar and RSS is the
220 preferred approach, even though the signal range differs from Bluetooth [26-27].
221 Besides methods for determining the location based on the measurement of radio signal strength
222 in the literature can be found a number of alternative solutions based on active measurements. These
223 methods may include, among others: the method of TOA (Time of Arrival) [28], as well as the method
224 2.4 GHz Phase Thistle. TOA method is based on measuring the time of radio waves propagation
225 between the transmitting device for which the location is determined and the receivers installed in
226 the Wi-Fi access points.
227 Times measurements obtained from multiple access points are further processed by an
228 algorithm performing the approximation of device distance monitoring. Even if the methods reach
229 reasonably high accuracy (location accuracy of 0.3 meters), it should be emphasized that this method
230 requires precise time synchronization of transmitting and receiving device, which cannot be obtained
231 without installing appropriate software into the device.
232 Despite the use of advanced computational algorithms in practice there are a number of
233 restrictions, which greatly affect the accuracy of positioning. Studies in [20] have shown that RSS
234 values, and thus the positioning accuracy, greatly depends on the orientation (rotation) of the
235 measuring device. This is due to irregularities in the radio signal, which measured power depends
236 on the direction of orientation of the antenna, component of reflected radio signal, and the proximity
237 of the user's body, which due to the high-water content in the human body absorbs a part of radio
238 signals [20], [29]. The abovementioned correlations denote that in practical applications the
239 measurements in phase of positioning almost always take place under different conditions from
240 measurements in the calibration phase. We must also take into account the fact that completely
241 different receiving equipment with different characteristics are usually used in the calibration and
242 positioning phases. Furthermore, in a range of practical applications there are variable propagation
243 conditions (e.g. a pallet of goods can significantly block the signal of the specified transceiver), and
244 there may be interference with other networks operating in the same frequency band. This all results
245 in errors in the specifications of localization indoors.
246 A very good example illustrating the technical level of existing solutions, may be the results of
247 a competition organized by Microsoft [30]. The main objective of the competition was to compare the
Sensors 2018, 18, x FOR PEER REVIEW 6 of 23

248 results obtained for the different positioning technologies inside buildings. 22 teams participated in
249 this competition, representing both the academic community and the world of industry. On the
250 technical side different approaches were represented from the use of a Wi-Fi network in the existing
251 premises, to the use of dedicated infrastructure e.g. Bluetooth, magnetic resonators or ultrasonic
252 transmitters.
253 The results of this competition can be regarded as a good benchmark for existing solutions. It
254 shows that the average positioning error, which was achieved by individual solutions ranged
255 between 0.72 m and 10.22 m. Only 3 teams were able to achieve an accuracy of less than 2 m while
256 the accuracy of 3 m was achieved by half of teams. In most cases, a significant increase of fault
257 positioning was recorded, in case when localization /orientation of the furniture changed. Another
258 important conclusion of the measurements is the fact that for a part of the points it was relatively
259 easy to determine locations (regardless of the solution), while positioning of the remainder of the
260 points were much more subject to error.
261 Summarizing the results of the competition, the organizers concluded that solutions for
262 positioning indoors, where the level of accuracy would allow for the implementation of a number of
263 desirable applications, for example such that would allow to locate products on a specific shelf, were
264 not yet available at the moment.
265 Another interesting commercial system has been presented by Zebra. This system provides an
266 accuracy of 0.5 meters and works with Wi-Fi, Bluetooth and technologies where the devices sends
267 beacons. The main drawbacks (and the reasons by which the solution has not reached the expected
268 success) are that Zebra’s system requires an active system, i.e., the users should install dedicated
269 software.
270 However, as important feature of that system is multi-technology. Because it is required that
271 proposed system allows to monitor / locate as many people moving in the monitored premises, it is
272 desirable to use for this purpose radio solutions implemented in different technologies that may be
273 available on devices owned by users. Those technologies should notably include: Wi-Fi, Bluetooth,
274 RFID, IEEE 802.15.4 networks. Simultaneous use of several technologies, on the one hand allows to
275 significantly increase the number of monitored devices (users), on the other hand, increases the
276 complexity and cost of the system.
277 Furthermore, due to the fact that the abovementioned technologies work in the same frequency
278 range-2.4 GHz, the problem becomes interference between the systems, which further results in less
279 positioning accuracy. With this situation, it is necessary to develop a high-accuracy system capable
280 of successfully working in multi-technology environment while maintaining low energy costs (low
281 operational expenditures).

282 2. Indoor positioning system


283 In this section we specify and offer implementation details of the indoor positioning system for
284 high positioning accuracy. The system revolves around two main ideas: (1) an enhanced multi-
285 technology multi-antenna hardware based on separation of network access and communication
286 layer, and (2) an advanced fingerprints algorithm, which puts together calibration and positioning
287 phases.

288 2.1. Multi-technology multi-antenna scenario based on separation of network access and communication with
289 devices
290 The proposed positioning system associated with people staying indoors under monitoring is
291 based on the measurement of the power of radio signal sent by the device of the monitored using
292 different technologies in the 2.4 GHz band, i.e. Wi-Fi, Bluetooth, RFID, IEEE 802.15.4 networks. It
293 should be emphasized that while there are systems for localizing objects inside rooms using more
294 than one technology, we are no aware of any solutions (based on passive measurements) providing
295 support for all the above-mentioned technologies. Current solutions typically use for positioning no
296 more than two wireless technologies, one of which is usually Wi-Fi technology and the second one,
297 according to the system may be, for example, IMU [22], Bluetooth [31] or the use of ultrasound [32].
Sensors 2018, 18, x FOR PEER REVIEW 7 of 23

298 Localization analysis based on measurements of sets of antennas show high positioning
299 efficiency in indoor scenarios but the most important inconvenience of using arrays of antennas is
300 the extra price that this hardware carries with. Therefore, we provided a multi-antenna system with
301 a single physical transceiver where its control is centralized and fully softwarized [33]. For providing
302 this system we used some of our previous ideas presented in [34-37].
303 The idea is to separate clearly the functionalities of network access and communication with
304 devices. On this way, the multiple protocols making part of the array module will include simple
305 hardware addressed to receive the beacons from the mobile devices and forward to the intelligence
306 of the system. No other action will be required from the transceiver, which reduces essentially its cost
307 and makes possible to populate the analyzed space with a high number of small and low functional
308 single-transceiver modules.
309 The intelligence of the system considers all the potential technologies and develop different
310 virtualized nodes for managing the identification of the devices (belonging to different technologies)
311 [38]. In our solution we develop a virtualized software platform containing a number of nodes, each
312 one managing a different technology.
313 In the multi-technology multi-antenna solution presented in this paper we introduced the most
314 important technologies in 2,4 GHz bandwidths, concretely we included Wi-Fi and technologies based
315 on IEEE 802.15.4 standard. Fig. 1 shows the functioning of multi-technology multi-antenna system.
316 The radio part will alternately carry out the device functions in abovementioned technologies (time
317 division will be used). According to [39], the co-existence of many radio technologies depends on
318 three factors: frequency, location in space and time. In this way, if the individual radio networks
319 differ at least on one of the above factors, they will be able to function properly. In our case, as we
320 use the same transceivers (which means the same frequency and location), the coexistence of multiple
321 technologies can be implemented only through the introduction of time-division. The two main
322 advantages of this solution are the lack of interference between devices of different technologies
323 (since the transceiver kit performs the function of only one technology at a certain interval of time),
324 and significantly lower cost of implementation in comparison with the parallel installations for each
325 technology. From the point of view of developed system, the most important advantage is the
326 possibility of providing more accurate measurements of the object. This is due to two factors: firstly,
327 greater accuracy is the result of a lack of interference arising from the co-existence in the given
328 environment of several radio technologies operating in the same frequency band; secondly,
329 significantly lower cost of the components of the radio network allows for deployment of a high
330 number of transceivers in an area, increasing the accuracy of user’s device positioning.
Virtual Machines

Demodulator
Modulator
Modulator
Demodulator
Demodulator
Modulator Demodulator

Sequence
Hypervisor

analizer and
dealer
NIC driver
NIC driver
NIC driver
NIC driver
NIC driver

Baseband chip sequence


(1Gbps coax cable)
Sensors 2018, 18, x FOR PEER REVIEW 8 of 23

331 Figure 1. Multi-technology multi-antenna functional architecture.

332 The Wi-Fi and 802.15.4 transceivers perform few functionalities related to signal demodulation
333 and transmission of the digital information to the central virtualized platform, which will address the
334 digital information to the appropriate virtual machine depending on the technology.
335 The Wi-Fi transceivers contain uniquely the Physical Medium Dependent (PMD) layer which is
336 responsible of converting the electrical signal in arrays of bits. It may use different encoding
337 techniques (among others, DBPSK, DQPSK, BPSK, QPSK, CCK, QAM) depending on the version of
338 the standard used and the distance of the mobile device) to the transceivers. The encoding technique
339 (and modulation) utilized decides the maximum bitrate in each channel and, in consequence, the
340 bitrate necessary to transport the information from the transceiver to the platform.
341 From the point of view of 802.15.4, the transceivers only manage the baseband chip sequence at
342 the physical layer. The chip sequence is a direct mapping of raw data into a set of ones and zeros that
343 are afterwards modulated by the electrical signal (half-sine pulse factor used in offset quadrature
344 phase-shift keying, O-QPSK). In our solution this sequence is directly sent from the transceiver to the
345 platform, where the sequence is converted to raw data. The baseband chip sequence contains the
346 same information as raw data but requires higher bitrate due to the high overhead introduced by the
347 physical layer, which limits the transmission capacity between transceiver and platform, while
348 reducing the complexity in the modules. In fact, the baseband chip sequence includes PHY Protocol
349 Data Unit (PPDU) for synchronizing the signal within the receptor as well as additional information
350 as, e.g., the length of the frame. The baseband chip sequence contains more than eight times more
351 bits than raw data since each 4 bits of raw data are encoded in 32 chip values and, in addition, PPDU
352 is added.
353 The platform contains higher layer functionalities of the considered technologies. In Wi-Fi
354 technology, the platform assumes the functionalities of the Physical Layer Convergence Procedure
355 (PLCP) responsible of analyzing the preamble and PHY headers, and the functionalities of layer 2
356 and layer 3. In fact, the transceiver continuously sends demodulated bits but it does not interpret the
357 sequence of bits. It is the platform which, upon receiving the bit sequence, analyzes it and extracts
358 the information related to the whole transmission (starting from the understanding of the bounds of
359 the transmission in the sequence of bits through the localization of the preamble).
360 The platform containing the management of the technologies is based on a virtualized machine,
361 where each Virtual Machine (VM) hosts one of the technologies. The hypervisor is in charge of
362 controlling the functioning of the VMs as well as managing the communication with the network
363 card. The frames arrived from the NIC (Network Interface Card) are copied in all the VMs irrespective
364 of the technology of the frame and the VM is in charge of discarding/accepting the frame if
365 belongs/does not belong to the technology of the VM. In all the technologies, the MAC layer has been
366 implemented, so that MAC layer takes decisions about discarding/accepting the frame.
367 In addition to MAC functionalities, each VM implements positioning modules. These modules
368 are responsible of providing the positioning process, which is technology-dependent (the fingerprints
369 are different for each one of the technologies). Let us remark that technology-dependence appears for
370 all the technologies, which makes the implementation easier.
371 For more details about the implementation of the platform, please see [33].
372 In conclusion, the proposed modulated system is based on softwarization of the main
373 communication (with devices) functionalities, which is possible thanks to the poor communication
374 (only based on beacons) framework needed in the indoor positioning system and allows for reducing
375 CAPEX and OPEX thanks to the simplicity of the radio system. This results in cost-saving multi-
376 antenna and flexible (softwarization) multi-technology system.

377 2.2. Fingerprints analysis


378 Current fingerprints algorithms assume a two-phase methodology for device positioning. At the
379 beginning the room is measured (off-line) in many points and a map with the measurements is the
380 basis of device positioning which is performed at a second stage of the process (on-line). The main
Sensors 2018, 18, x FOR PEER REVIEW 9 of 23

381 disadvantage of such an approach is that the conditions of the system in the two stages (off-line and
382 on-line) normally are not the same; a clear example is the introduction of obstacles (e.g., other
383 customers) in the on-line phase of the process which were not present in the map building.
384 Therefore, we propose a step forward in fingerprints methodology that assumes that map
385 measurements are performed in the same conditions as final device positioning. The idea is to create
386 a map based on statistics obtained from a high number of RSS measurements taken while the shop is
387 open.
388 For this scope, the classical fingerprints mathematical tools are no longer valid since such
389 algorithms do not consider uncertainties due to real working conditions.
390 In our approach, the calibration process is separated in two different phases. The first one is
391 processed “off-line”, i.e., the calibration measurements are obtained when no customers are in the
392 analyzed space. These measurements are very precise and the process is slow due to the closeness of
393 the testing points. The closeness of testing devices is a key issue for achieving high positioning
394 accuracy and it is limited by the sensibility of the antennas and other hardware. The second phase of
395 the calibration is performed together with the positioning process. This is, the calibration
396 measurements are obtained when customers are in the analyzed space (“on-line” calibration). Several
397 testing devices are located in the room and operate jointly to customers’ devices, which provoke
398 interferences, echoes, obstacle effects, etc. The on-line calibration measurements are interpolated with
399 the off-line ones in order to obtain the final fingerprints map.
400 In the customers’ devices positioning process, the devices are localized in given point of the
401 analyzed space by comparing the signal arrived from the device with the fingerprints map. Statistical
402 tools are used for obtaining the position with highest accuracy.
403 Off-line Calibration
404 The fingerprints estimation consists of taking statistically significant RSS measurements from all
405 the testing points within the analyzed area. The minimum distance between two neighboring testing
406 points is lower-bounded by the positioning accuracy required by the customer of the service (i.e., a
407 shop where the system is installed) as well as by the limits of hardware and statistical tools. A
408 discussion of the minimum distance may be found in the next sub-section. By the moment, let us
409 assume that, in off-line calibration, we have P testing points around the analyzed surface.
410 The testing device sends an identifiable data sample and the RSS of the signal arrived to the
411 antennas is measured in each one of the N transceivers. The testing device repeats this operation in
412 order to obtain a number (say M) of RSS measurements in the transceivers, such that a consistent
413 average may be obtained. The off-line calibration map contains the N-dimensional mean values (of
414 M measurements) for each testing point, together with the physical location of the testing point.
415 The off-line calibration corresponds to the classical calibration of fingerprints methods.
416 On-line Calibration
417 This process runs on parallel to the users’ devices positioning process, i.e., during the normal
418 activity of the shop.
419 Q testing devices are located widespread in the analyzed area, so that the on-line calibration
420 counts with Q testing points, which in number are less than the P testing points in off-line calibration.
421 Each one of the Q testing devices sends identifiable data samples in intervals of time τ. For the scope
422 of the algorithm, fingerprints values are taken only based on the most recent M samples whereas the
423 previous samples are discarded. This signal is received in all the N transceivers and the RSS is
424 measured.

425 Let 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 be the RSS value of the signal emitted by the testing device situated in → (q=1..Q)
𝑥𝑞

426 in time equal to current timer-(M- m)•τ, (m=1..M), and measured in transceiver n (n=1..N).
427 Due to changing conditions of the environment, the last (in time) calibration measurements are
428 likely to be more precise than previous ones since the former have been taken in conditions that are
429 much more similar to the current conditions than previous measurements. Therefore, we introduce
Sensors 2018, 18, x FOR PEER REVIEW 10 of 23

430 a moving average of the calibrated measurements for becoming the reference measurement for
431 localizing customers’ devices.
432 We consider Weighted Moving Average, WMA, and Exponential Moving Average, EMA, of the
433 calibrated measurements. The suitability of the moving average depends on the scenario, and
434 specifically, on the measuring interval τ, as it will be shown further on in the results.
̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅

435 Let [𝑊𝑀𝐴𝑥𝑞 ] and [𝐸𝑀𝐴𝑥𝑞 ] be the vectors of Weighted Moving Averages and Exponential
436 Moving Averages of the RSS measurements arrived from testing device located at →, as follows:
𝑥𝑞
𝑀

∑𝑀−1
𝑚=0(𝑀 − 𝑚) × 𝑅𝑆𝑆 (𝑀−𝑚),1
𝑥𝑞
∝× ∑(1−∝)𝑀−𝑚 × 𝑅𝑆𝑆 𝑥𝑞 𝑚1

𝑀 × (𝑀 + 1)/2 𝑚=1
→ 𝑀
∑𝑀−1
𝑚=0(𝑀 − 𝑚) × 𝑅𝑆𝑆 (𝑀−𝑚),2
𝑥𝑞 →
̅̅̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅̅ 𝑀−𝑚
[𝐸𝑀𝐴𝑥𝑞 ] = ∝× ∑(1−∝) × 𝑅𝑆𝑆 𝑥𝑞 𝑚2 ,
→ →
[𝑊𝑀𝐴𝑥𝑞 ] = 𝑀 × (𝑀 + 1)/2 ∝= 1 − 0.05𝑀 (1)
𝑚=1
⋮ ⋮
→ 𝑀
∑𝑚=0(𝑀 − 𝑚) × 𝑅𝑆𝑆 𝑥𝑞 (𝑀−𝑚),𝑁
𝑀−1

∝× ∑(1−∝)𝑀−𝑚 × 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑁
[ 𝑀 × (𝑀 + 1)/2 ] [ ]
𝑚=1

437 The value ∝= 1 − 0.05𝑀 ensures that the sum of the weights of the M last samples equals 95%
438 of all the measurement samples record.
439 With the RSS values of the on-line testing points, we may estimate the (RSS) values of the other
440 (P-Q) testing points of the investigated area (these P-Q testing points were presented in the off-line
→ ∗
441 calibration but not in the on-line). For estimating the RSS values of the → , 𝑝 = 1. . 𝑃 (say 𝑅𝑆𝑆 𝑥𝑝 ), we
𝑥𝑝

442 scale the off-line fingerprints values 𝑅𝑆𝑆 𝑥𝑝 by observing the scaling tendency of the nearest
443 (physically) on-line testing points. This methodology assumes that close testing points behave
444 similarly. For the scope of estimation, we perform interpolation of the four nearest on-line testing
→ ∗
445 points. Concretely, the current measurement 𝑅𝑆𝑆 𝑥𝑝 for off-line testing point → , which had an off-
𝑥𝑝

446 line calibration RSS value equal to 𝑅𝑆𝑆 𝑥𝑝
is calculated with the interpolation formula:
→ → → →
𝑅𝑆𝑆 𝑥𝑝 × ̅̅̅̅̅̅̅
4 → →
→ ∗ ∑4𝑗=1 𝑑 𝑥𝑝 𝑥𝑗 − 𝑑 𝑥𝑝 𝑥𝑖 𝑀𝐴𝑥𝑖
𝑅𝑆𝑆 𝑥𝑝
= ∑[ → → × → ] (2)
𝑖=1 3 × ∑4𝑗=1 𝑑 𝑥𝑝 𝑥𝑗 𝑅𝑆𝑆 𝑥𝑖

̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅

447 , where 𝑀𝐴𝑥𝑖 is any of the moving averages vectors ([𝑊𝑀𝐴𝑥𝑖 ] or [𝐸𝑀𝐴𝑥𝑖 ]) of the on-line testing


448 points (→) that are closest to the estimated off-line testing point (→ ); 𝑅𝑆𝑆 𝑥𝑖 are the four off-line RSS
𝑥𝑖 𝑥𝑝

→ →
449 values of the testing points (→) closest to the estimated point (→ ); 𝑑 𝑥𝑝𝑥𝑖 is the Euclidean physical
𝑥𝑖 𝑥𝑝

450 distance (in centimeters) between the testing point to be estimated ( → ) and the closest on-line testing
𝑥𝑝

451 points (→).


𝑥𝑖

452 Within the above operations, we obtain a full fingerprints map, which is measured and
453 estimated in real operative conditions. The map considers interferences, signal shadowing, and other
454 effects caused by the customers’ devices activity.
455 The estimation process requires high processing power since large amount of data is processed
456 in each interval of time, τ. The same will occur with the localization process, however, the latter will
457 be run only when customer’s device signal is received in the transceivers.
458 Device positioning
Sensors 2018, 18, x FOR PEER REVIEW 11 of 23

459 Once we have the fingerprints, the customer’s device can be positioned in the analyzed space by
460 comparing its signal RSS (arrived to the transceivers) with the calibrated fingerprints map. Let [RSSd]
461 be the N-vector of RSS values of the signal arrived from the user’s device, d.
462 In order to estimate the location of the customer’s device, we calculate the Mahalanobis distance,
̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅

463 𝑑𝑀𝑝𝑑 , between [RSSd] and each one of the P vectors [𝑊𝑀𝐴𝑥𝑝 ] or [𝐸𝑀𝐴𝑥𝑝 ] , 𝑝 = 1. . 𝑃 obtained from

464 the P testing points, as indicated in (3). The location of the customer’s device will be the same than
465 the testing device p’, p’Є[1,P], whose Mahalanobis distance 𝑑𝑀𝑝′𝑑 is minor. The Mahalanobis
466 distance is an statistical tool capable of distinguishing variance of the different measurement points
467 (transceivers), because those with high variability of the RSS measurements are considered to be
468 exposed to uncertainty, that is to say, this distance also takes into account which transceivers provide
469 more (secure) information for localizing a given point.
470 The Mahalanobis distance is calculated as follows:
̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅

𝑑𝑀𝑝𝑑 = √([𝑅𝑆𝑆 𝑑 ] − 𝑀𝐴𝑥𝑝 )𝑇 × [𝐶𝑜𝑉 𝑥𝑝 ]−1 × ([𝑅𝑆𝑆 𝑑 ] − 𝑀𝐴𝑥𝑝 ) (3)
̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅
→ →
471 , where 𝑀𝐴𝑥𝑝 is any of the moving average vectors: [𝑊𝑀𝐴𝑥𝑝 ] or [𝐸𝑀𝐴𝑥𝑝 ], and [𝐶𝑜𝑉 𝑥𝑝 ] is the co-

472 variance matrix as indicated in (4). In order to calculate [𝐶𝑜𝑉 𝑥𝑝 ], the system should store the last M
473 RSS values of all P testing points.

[𝐶𝑜𝑉 𝑥𝑝 ]
𝑀 𝑀 𝑀
1 → ̅̅̅̅̅̅̅
→ 1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅
→ 1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅

× ∑(𝑅𝑆𝑆 𝑥𝑝 𝑚1 − 𝑀𝐴𝑥𝑝 [1])2 × ∑(𝑅𝑆𝑆 𝑥𝑝 𝑚1 − 𝑀𝐴𝑥𝑝 [1]) × (𝑅𝑆𝑆 𝑥𝑝 𝑚2 − 𝑀𝐴𝑥𝑝 [2]) … × ∑ (𝑅𝑆𝑆 𝑥𝑝 𝑚1 − 𝑀𝐴𝑥𝑝 [1]) × (𝑅𝑆𝑆 𝑥𝑝 𝑚𝑁 − 𝑀𝐴𝑥𝑝 [𝑁])
𝑀−1 𝑀−1 𝑀−1
𝑚=1 𝑚=1 𝑚=1
𝑀 𝑀 𝑀
1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅
→ 1 → ̅̅̅̅̅̅̅

2
1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅

(4)
= 𝑀 − 1 × ∑(𝑅𝑆𝑆 𝑚2 − 𝑀𝐴 [2]) × (𝑅𝑆𝑆 𝑚1 − 𝑀𝐴 [1]) 𝑀 − 1 × ∑ (𝑅𝑆𝑆 𝑚2 − 𝑀𝐴 [2]) … × ∑ (𝑅𝑆𝑆 𝑥𝑝 𝑚2 − 𝑀𝐴𝑥𝑝 [2]) × (𝑅𝑆𝑆 𝑥𝑝 𝑚𝑁 − 𝑀𝐴𝑥𝑝 [𝑁])
𝑥𝑝 𝑥𝑝 𝑥𝑝 𝑥𝑝 𝑥𝑝 𝑥𝑝
𝑀−1
𝑚=1 𝑚=1 𝑚=1

𝑀 𝑀 𝑀
1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅
→ 1 → ̅̅̅̅̅̅̅
→ → ̅̅̅̅̅̅̅
→ 1 → ̅̅̅̅̅̅̅

× ∑(𝑅𝑆𝑆 𝑥𝑝 𝑚𝑁 − 𝑀𝐴𝑥𝑝 [𝑁]) × (𝑅𝑆𝑆 𝑥𝑝 𝑚1 − 𝑀𝐴𝑥𝑝 [1]) × ∑ (𝑅𝑆𝑆 𝑥𝑝 𝑚𝑁 − 𝑀𝐴𝑥𝑝 [𝑁]) × (𝑅𝑆𝑆 𝑥𝑝 𝑚2 − 𝑀𝐴𝑥𝑝 [2]) … × ∑ (𝑅𝑆𝑆 𝑥𝑝 𝑚𝑁 − 𝑀𝐴𝑥𝑝 [𝑁])2
[ 𝑀 − 1 𝑀 − 1 𝑀−1 ]
𝑚=1 𝑚=1 𝑚=1

̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅

474 , being 𝑀𝐴𝑥𝑝 [𝑖] the i-element of any of the moving average vectors: [𝑊𝑀𝐴𝑥𝑝 ] or [𝐸𝑀𝐴𝑥𝑝 ].

475 The co-variance matrix computes the correlation between the RSS-samples measured in the
476 transceivers, so that the matrix shows how uniform are the samples arrived to each module.

477 2.3. Discussions on assumptions

478 The system efficiency is bounded by both limitations of hardware and software as well as
479 variability of environment conditions. The hardware presents limitations on measurement accuracy,
480 such limitations influence the minimum physical distance that can be distinguishable by the system,
481 so that the RSS measurement of devices located closer than a given threshold are not differentiated.
482 As a result, the localization of devices with a higher precision is not possible. Regarding to the
483 software, even if the system processes a considerable amount of data, the tests have not shown issues
484 on data processing, so we assume that the implementation of the software is good enough to process
485 the data arrived from testing points and customers’ devices.
486 At last, high environment variability could require a high rate of measurements of on-line
487 calibration in order to constantly update the calibration map. Therefore, the system should have the
488 capability to understand the environment variability and try to adjust the sampling rate of probes
489 during the on-line calibration.
490 Hardware limitations causing minimum distance between testing points and frequency of on-
491 line calibration probing will be described below.
Sensors 2018, 18, x FOR PEER REVIEW 12 of 23

492 2.3.1. Distance between testing points

493 The precision of the off-line calibration map depends on the transceivers’ ability to distinguish
494 two testing devices which are located close from each other. The scope of the transceivers is limited
495 by the physical distribution of the analyzed space, hardware limitations, power-consumption
496 limitations (power of the signal) and others. In general, we need to assume a trade-off between
497 accuracy and costs since better accuracy may be achieved only by increasing costs (better hardware,
498 more powerful signal, longer calibration process due to a higher number of testing points, etc.). The
499 other bound to the precision is the requirement of the service contractor, i.e., it may be that the
500 proprietary of the shop needs a lower positioning accuracy than the one allowed by the system. In
501 this case, it has not sense to invest in a very precise calibration map.
502 In the case that the accuracy is limited by the system capabilities, then we need to understand
503 which is the minimum distance between testing points (accuracy of calibration map) that can be
504 properly measured by the transceivers. For this scope, we should analyze the RSS (measured in the
505 transceivers) of the signals coming from two neighboring testing devices and check whether they are
506 statistically different. In this context, the first step of the off-line calibration is to measure the RSS of

507 the signal arrived from a testing device situated in two (physically) close points → and →. The test
𝑥𝑝 𝑥𝑞

→ →
508 should be repeated several times (say M) for each one of the points. Let 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 and 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 ,

509 m=1..M, n=1..N, be the RSS measurements from testing points → and →, respectively.
𝑥𝑝 𝑥𝑞

510 We consider that → and → are statistically distinguishable if any average or variance of
𝑥𝑝 𝑥𝑞

→ →
511 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 and 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 are different. The reason for this is that, during the device positioning
512 process the system will make use of the Mahalanobis distance, which considers both average and
513 variance (note that the Mahalanobis distance between the measurements of the user’s device and the
514 two calibration points, 𝑑𝑀𝑝𝑑 and 𝑑𝑀𝑞𝑑 will be different if average or variance is different). Thus, it

515 is possible that the customer’s device is properly localized even if the averages of 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 and

516 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 are equal while maintaining different only the variances. Fig. 2 shows an example of
→ →
517 scenario, where the blue and red points are the averages of 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 and 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 , respectively,
518 whereas the ellipses represent the increasing Mahalanobis distance from the averages. If a device is
519 located in a given point (whose RSS measurement is represented as a black point in the figure), then

520 the device should be clearly localized in → despite that the RSS is closer to the average of the testing
𝑥𝑞

521 point →.
𝑥𝑝
Sensors 2018, 18, x FOR PEER REVIEW 13 of 23

Antenna 2
Antenna 1
522 Figure 2. Example of Mahalanobis distance for two neighboring points.

523 For calculating whether → and → are statistically different. the first step is to compare the
𝑥𝑝 𝑥𝑞

524 variances of the (→ and →) testing measurements. The variance is compared separately for each one
𝑥𝑝 𝑥𝑞

525 of the N (assumed normal) measurement sets, i.e., we will consider the two testing points being

526 statistically different if the RSS samples measured in any transceiver coming from → and → are not
𝑥𝑝 𝑥𝑞

527 equal for a given significance level (90%). We use the Cochran test and compare with the table values
528 for a 90%-significance level.
529 In the case that all the variances are equal, then we compare the averages by using the MANOVA
530 test. The MANOVA uses any of the statistics existing for comparing the means; in our case we will
531 use the Hotelling's T-squared statistic and compare with the F-distribution values, which permits to
532 reject (or not) the null hypothesis that the averages are equal (at a required significance level). We
533 selected the Hotelling’s T2 statistic because it describes the behavior of the Mahalanobis distance

534 (T2=M•𝑑𝑀2 → → ) between the averages of the two groups ( → and →) and compare this distance with
𝑥𝑝 𝑥𝑞 𝑥𝑝 𝑥𝑞

535 tabulated values. Previously was mentioned why Mahalanobis distance describes better the
536 incidence of the transceivers into the measurements.
537 Let us remark that the statistical tools used here require many data in order to compare variances
538 and averages, which may be obtained only through a long testing process. Here a balance between
539 the necessary positioning accuracy and testing process costs should be acquired. In our tests the
540 results showed that a room of 100 m2 scanned by six transceivers may obtain a minimum distance of
541 40 cm.

542 2.3.2. On-line calibration probing frequency

543 Whereas the off-line calibration measures the effect of stable conditions (furniture, transceivers,
544 etc.) into the calibration map, the on-line calibration estimates the changing-conditions effects. These
545 latter are more difficult to estimate and an adequate measurement frequency during on-line
546 calibration is crucial in order to properly calculate the calibration map: too rare measurements could
547 provoke that older measurements provide false information about the state of the analyzed space
548 (e.g., old measurements provide information about the effect of other customers’ devices, which are
549 not present anymore in the room) and addresses a false calibration map; on the other hand, too often
Sensors 2018, 18, x FOR PEER REVIEW 14 of 23

550 measurements could make the measurements uniform, which would rescind important information
551 related to the reliability of the transceivers (stating which one is more precise).
552 Therefore, we introduce a mechanism for adjusting the sampling rate of the on-line calibration
553 measurements, such that the probing is not too recurrent neither too sporadic. Let us consider the M
554 last on-line calibration measurements, which shape a time series. The objective is that the time series
555 is not a purely stationary process (caused by same conditions of the environment during all the
556 measuring period of the time series, i.e., the measurements are too frequent and all of them give the
557 same information), and, at the same time, it contains a non-jerky deterministic component (jerky
558 series is caused by suddenly changing conditions of the environment, which would indicate that old
559 measurements are not showing the current state of the environment anymore). Thus, a non-stationary
560 with non-jerky deterministic component series ensures that the time series picks up changes
561 experienced into the environment while avoiding sudden changes.
562 We decompose the time series in K periods (K groups of samples), such that each period contains
563 M/K measurement samples. The value of K should be as high as possible while assuring that M/K is
564 high enough to obtain the expected values (mean and variance) with enough statistical significance.
565 We first check that the series is not stationary-like and, afterwards, that it is not jerky:
566 1- We consider that the time series has a stationary behavior if each one of the K groups have
567 the same average. For this, we apply the ANOVA method, which makes use of the variance of the
568 groups in order to understand whether the differences of the averages are caused by the variations
569 inside the group or by differences between groups. In the case that the averages differ significantly
570 (90% significance level, i.e., the value in the F-distribution is F0.9,K-1,M-K), then the probing frequency is
571 maintained, otherwise the probing frequency is reduced in a 10% of the current value.
572 2- The series is jerky when there are appreciable differences between consecutive groups of
573 samples. Therefore, in order to investigate the jerky character of the time series, we first calculate the
𝑀
̅̅̅̅̅̅̅̅̅̅̅
→ 𝐾 (𝑘+1)× →
574 mean values of each group of samples = 𝑅𝑆𝑆 𝑥𝑝 𝑘,𝑛 = ×∑ 𝑀
𝐾
𝑅𝑆𝑆 𝑥𝑝 𝑖,𝑛 , 𝑘 = 1. . 𝐾 − 1, 𝑛 = 1. . 𝑁,
𝑀 𝑖=𝑘× +1
𝐾

̅̅̅̅̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅̅

575 and afterwards we calculate the values |𝑅𝑆𝑆 𝑥𝑝 𝑘+1,𝑛 − 𝑅𝑆𝑆 𝑥𝑝 𝑘,𝑛 | , 𝑘 = 1. . 𝐾 − 1, 𝑛 = 1. . 𝑁 , and

→ →
|𝑅𝑆𝑆 𝑥𝑝 𝑖,𝑛−𝑅𝑆𝑆 𝑥𝑝 𝑖−1,𝑛 |
576 compare each value with the average ∑𝑀
𝑖=2 , n=1..N. Notice that we compare the
𝑀−1

577 jumps between the groups with the mean jump value between consecutive samples in the whole
578 series. This is because we are searching a jerky behavior between groups and not between single
579 samples, but the differences between samples offer information about the “acceptable” values of the
̅̅̅̅̅̅̅̅̅̅̅

580 jumps. In the case that any value 𝑅𝑆𝑆 𝑥𝑝 𝑘,𝑛 , 𝑘 = 1. . 𝐾 − 1, 𝑛 = 1. . 𝑁 fulfills formula (5) for any of the

581 N transceivers, then the sampling rate is increased in a 10%. Otherwise, is maintained without
582 modifications.
→ →
𝑥𝑝 𝑥𝑝
̅̅̅̅̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅̅
→ |𝑅𝑆𝑆 𝑖,𝑛−𝑅𝑆𝑆 𝑖−1,𝑛 |
583 |𝑅𝑆𝑆 𝑥𝑝 𝑘+1,𝑛 − 𝑅𝑆𝑆 𝑥𝑝 𝑘,𝑛 | > 𝑈 × ∑𝑀
𝑖=2 (5)
𝑀−1

584
585 Let us remark that the value U is a set-up parameter that depends on the analyzed space. In
586 order to calculate proper U value, an experiment should be provided at the beginning of the system

587 lifecycle: a testing device situated in → should send signals in stable (stationary) and changing
𝑥𝑝
Sensors 2018, 18, x FOR PEER REVIEW 15 of 23

588 conditions. The value of U should be such that allows to understand whether the value
̅̅̅̅̅̅̅̅̅̅̅̅̅̅
→ ̅̅̅̅̅̅̅̅̅̅̅

589 |𝑅𝑆𝑆 𝑥𝑝 𝑘+1,𝑛 − 𝑅𝑆𝑆 𝑥𝑝 𝑘,𝑛 | is due to stationarity or to changing conditions of the environment.

590 The mechanism for sampling rate checking and adjustment is run when M new probes are
591 measured in the transceivers. In this way the operations described previously are based always on

592 different 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 values. Sampling rate adjustment is not required to be a real-time process, so
593 adjusting the period in M•τ time intervals looks quite appropriate. Adjustment of the sampling rate
594 requires a communication framework between the system and the testing devices in order to inform
595 about the period’s value, however, this capability has not been developed yet in our system (future
596 work). Sampling rate adjustment allows energy-saving since the testing devices situated in → , 𝑝 =
𝑥𝑝
597 1. . 𝑃 may go to sleep during τ, such that they operate only when it is necessary. Once again, this
598 functionality has not been developed in our current implementation.

599 3. Testbed experiments


600 We set the tests to the following parameters: 1) only one customer’s device is tracked, its position
601 is well-known, so that the difference between real position and estimated position (provided by the
602 system) can be calculated. The localization of one unique customer’s device is only for the scope of
603 these tests inasmuch as the system may localize many devices on parallel, however, it would be very
604 difficult to show system efficiency in tests measuring mobile external devices’ signal (without
605 knowledge of exact devices’ situation); 2) even though the modules are multi-technology, for the aim
606 of these tests, the customer’s device communicates uniquely through Bluetooth technology. Some
607 tests are provided during the normal operation of the system, i.e., many customers with devices
608 communicating through several technologies are in the shop and their devices introduce
609 interferences, signal shadows and obstacles.
610 The system has been implemented and tested in indoor (100-square meters 400-cubic meters
611 furnished room) scenario. The tests include functional tests and performance tests: in the functional
612 tests we analyze the parameters of the model (calibration phases) and, concretely, the minimum
613 distance between testing points, the off-line accuracy in multi-antenna environment, the on-line
614 calibration setting-up parameters and the impact of the moving average into the results of the
615 algorithm. In the performance tests we compare the functioning of our system with the same system
616 that does not include on-line calibration phase. This test allows to compare the system with other
617 solutions proposed until now.

618 3.1. Off-line calibration - Minimum distance between testing points

619 This test aims to provide the minimum distance between testing points, such that the system
620 may distinguish the signals arriving from neighboring testing points. The minimum distance shows
621 the points of the calibration map. This distance is not necessarily the precision of the system but it is
622 the precision bound. In other words, even if the calibration map has points, which are separated by
623 the minimum distance, other factors such as obstacles, interferences, etc., may cause that localization
624 of devices in accurate position of the map is not reachable. Let us remark that this minimum distance
625 is related with the value obtained in the tests provided in [30], so this distance may be used for
626 comparison (taking into account the differences in the scenario) with the solution proposed there.
627 The scenario of the tests is as follows: The shop contains 6 transceivers located in the perimeter
628 of the room. The customer’s device is located in the center of the room and the RSS is measured 100

629 times in the transceivers (𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 𝑚 = 1. .100, 𝑛 = 1. .6). Then the customer’s device is located in the

630 horizontal axis (west-east direction of the room) in distances equal to 0.2, 0.4, 0.6, 0.8 and 1.0 m of →
𝑥𝑝
Sensors 2018, 18, x FOR PEER REVIEW 16 of 23


631 and the RSS is measured 100 times in all the transceivers ( 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 , → = → + 𝑥, 𝑥 =
𝑥𝑞 𝑥𝑝
→ →
632 [0.2,0.4,0.6,0.8,1.0], 𝑚 = 1. .100, 𝑛 = 1. .6). The 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 values are compared with 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 through
633 Cochran and MANOVA tests. The Cochran test (Cochran’s C test) showed that the variances of the
634 measurements in all the transceivers were similar since the C values obtained from the measurements
635 were much higher than the 0.23 critical value obtained from the tables (6 variables and 100 samples),
636 so homoscedasticity is confirmed.

637 Therefore, we used MANOVA for comparing the mean of the 𝑅𝑆𝑆 𝑥𝑝 𝑚𝑛 values with each one of

638 the means of 𝑅𝑆𝑆 𝑥𝑞 𝑚𝑛 values. The Hotelling's T2 value of the test was compared with F-distribution
639 table (F=2.72 for 6-numerator and 193-denominator freedom degrees and 0.9-significance level) for
640 each one of the testing points →. The results are presented in Fig. 3.
𝑥𝑞

12

10
10,1
8
8,01
T2/2

4 4,45
3,14 F=2,72
2 1,64
0
0 0,2 0,4 0,6 0,8 1
Distance between xp and xq [m]

641 Figure 3. MANOVA results for increasing distance between → and →.


𝒙𝒑 𝒙𝒒

642 As we can see, for a significance interval equal to 0.9, the minimum distance between testing

643 points is 0.4 m. For distance equal to 0.2 m, the RSS measured from devices located in → and → are
𝑥𝑝 𝑥𝑞

644 not distinguishable, so the system is not capable to localize a device in one of these two points.
645 There are some considerations to do about these results. As mentioned above 0.4 m is not the
646 precision of the system since there are other factors which may reduce the precision. An example of

647 this is one sample in our test, which measured the device located in → + 0.4 𝑚. The RSS values
𝑥𝑝


648 measured (in all the transceivers) were very similar to the mean 𝑅𝑆𝑆 𝑥𝑝 value, which indicates that
649 in regards to that measurement, the positioning of the device would be erroneous. For all other
650 measurements, the positioning would be correct, however, this case shows that there exist other
651 factors limiting the efficiency of the system.

652 Once the minimum distance was defined to 0.4 m, we performed similar tests by locating the
653 testing device in the points of a 0.4-squared grid and provided the off-line calibration map. This map
654 has been used in further tests.

655 3.2. Off-line positioning efficiency

656 The following test aims to show the effect of the number of transceivers into the system
657 positioning efficiency. Concretely, we situated the testing device in many (500) known-location
Sensors 2018, 18, x FOR PEER REVIEW 17 of 23

658 points of the room and took measurements of the signal in N (N=3,4,5 and 6 transceivers). Based on
659 these measurements, the system took a localization decision by comparing the measurements with
660 the appropriate off-line calibration map (4 calibrations maps were built from off-line measurements
661 taken by 3,4,5 and 6 transceivers).
662 The difference between position and estimation (absolute value of 500 measurements) with
663 different number of transceivers is presented in Fig. 4. The figure shows the mean positioning
664 accuracy of the measurements on the 0.95-percentile (the positioning error upper bound of the 95%
665 most precise measurements).

666 Figure 4. Positioning accuracy vs. number of transceivers.

667 The system localizes the device in one of the points of the 0.4-square grid calibration map,
668 so the decision should be considered as correct if the selected point into the map is the closest to
669 the real location of the device. Therefore, differences lower than 0.2 m indicate a correct decision,
670 however, differences higher than 0.2√2 are always incorrect. Between 0.2 and 0.2√2 , the
671 correctness of the decision depends on the position of the device (i.e., if the localized position is
672 the closest map point).
673 In the case of 4 transceivers measuring the signals, the mean value of the differences
674 between localization and device location is lower than 0.2, however, there are many mistaken
675 decisions, which may be observed by the 0.95-percentile value equal to 0.32 > 0.2√2. Only for
676 the case of five and six transceivers measuring the RSS, the results show 100% correct
677 positioning decisions. High efficiency of the localization decision algorithm is explained thanks
678 to the fact that there is only one device in the room, so no interferences are present and, in
679 addition to this, the furniture of the shop has not been changed from the off-line calibration, so
680 the map perfectly fits the actual test scenario.
681 The variance of the results also diminishes with increasing number of transceivers. This is
682 produced by the more accurate behavior of the positioning algorithm.
683 The conclusion of these tests is that a high number of transceivers in a small room is needed
684 to provide high efficiency positioning system, which is possible only in the case of not expensive
685 transceivers. Such thing may be achieved with the hardware developed in our system.
686 3.3. On-line frequency analysis
Sensors 2018, 18, x FOR PEER REVIEW 18 of 23

687 In this test, the on-line sampling rate is analyzed during normal functioning of the system, i.e.,
688 when customers are in the shop. The customers create interferences, echoes and signal shadows that
689 introduce variability in the measurements.
690 In this scenario the transceivers receive many signals but, after analyzing them, only the testing
691 device’s signal is processed forward by the system. The emission period of the testing device is
692 manually changed in each experiment. The number of signal samples (emitted by a testing device)
693 measured by six transceivers were 200 samples (M=200).
694 The on-line sampling rate analysis is based on dividing the 200 samples in 4 groups (K=4), and
695 comparing the groups in order to understand whether the samples have stationary or jerky behavior.
696 For this scope, we calculate Fexp as the relation of the variance between groups and the mean
697 variance inside the group. This value is compared with the F-distribution for K-1 (3 in our case) and
698 M-K (196 in our case) degrees of freedom and 0.9-significance level. In the case that Fexp< F0.9,3,196,
699 then we may conclude that the series has a stationary behavior and the system should reduce the on-
700 line calibration sampling rate in order to obtain more significant measurements.
701 The tests have been repeated 10 times in order to obtain the confidence intervals of Fexp.
702 Changes between tests are subject to the variability of the environment; nevertheless, taking into
703 account that all the tests (for each tested rate) were performed in a short interval of time, we may
704 assume similar environment conditions, so the results tend to be Gaussian.
705 The results and the decision of the algorithm for different sampling rates are shown in the first
706 four columns of Table 1. The other columns of the table show the analysis performed for checking
707 whether the system should increase the sampling rate. For this, we compare the jumps between
708 consecutive groups with the mean difference between consecutive single samples, see formula (5).
709 From this comparison, the system concludes whether to increase the probing sampling rate. By
710 changing the value U, we may tune the minimum value of the sampling rate.

711 Table 1. Values for on-line calibration sampling rate testing.

On-line probing ̅̅̅̅̅̅̅̅̅̅̅̅̅̅


→ → →
𝑴
Sampling 𝒎𝒂𝒙 |𝑹𝑺𝑺𝒙𝒑 𝒒+𝟏,𝒏 − |𝑹𝑺𝑺𝒙𝒑 𝒊,𝒏 − 𝑹𝑺𝑺𝒙𝒑 𝒊−𝟏,𝒏 | Sampling
sampling Rate Fexp F0.9,3,196 ̅̅̅̅̅̅̅̅̅̅̅
→ 𝟑×∑
decrease 𝑹𝑺𝑺𝒙𝒑 𝒒,𝒏 | [dbm] 𝑴−𝟏 increase
𝒊=𝟐
[samples/min.]

40 9.40±2.21 5.14 No 3.31 2.26 Yes


60 8.80±2.01 5.14 No 2.28 2.27 Yes
80 7.86±4.12 5.14 No 2.24 2.25 No
100 7.01±4.51 5.14 No 2.17 2.22 No
120 6.24±1.32 5.14 No 2.15 2.19 No
140 5.24±1.39 5.14 No 2.14 2.20 No
160 5.18±3.88 5.14 No 2.14 2.16 No
180 4.74±1.17 5.14 Yes 2.12 2.15 No
200 4.01±1.00 5.14 Yes 2.13 2.12No
712 As we may observe in Table 1, for the current environment conditions, the on-line calibration
713 sampling rate adjustment algorithm will maintain a rate between 80 and 160 samples/min.
714 The table shows that, for some tests, the variance is clearly higher than for others. It is difficult
715 to understand this effect, specially if we take into account that this effect is repeated in several tests.
716 The unique supposition is that the variability of the environment has a high impact into the RSS
717 measurements.
718 3.4. Analysis of Moving Average

719 The selection of one of the two moving averages: Weighted Moving Average (WMA) or
720 Exponential Moving Average (EMA) has repercussions on the results of the positioning system.
721 Particularly, EMA gives higher weight to the last on-line calibration measurements than WMA,
Sensors 2018, 18, x FOR PEER REVIEW 19 of 23

722 which shares the importance of the measurements proportionally to their seniority. The selection of
723 EMA and WMA is closely related with the variability of environment conditions, in the same way as
724 the on-line calibration probing frequency: EMA avoids oscillations in results in the case of high
725 variable environment conditions, however, it fails on providing information about the reliability of
726 the transceivers.
727 Our analysis aims to compare the system efficiency for different on-line calibration measuring
728 frequencies and, specifically, for the values of frequency between 80 and 160 samples/min. The
729 calibration map during these tests does not change, so it is calculated off-line shortly before the test
730 is run. Changing the calibration map would cause difficulty for understanding the differences in test
731 results for different experiments.
732 The next figure shows the accuracy of the positioning algorithm for both moving averages and
733 increasing measuring sampling rate. The accuracy is calculated as the difference between location
734 and positioning (average and 0.95-percentile). The system works in normal conditions (customers in
735 the room) and the algorithm localizes the testing device at least 100 times in each experiment
736 (different positions) with six transceivers.

737 Figure 5. Positioning accuracy vs. on-line calibration sampling rate (EMA-based and WMA-based
738 algorithms).

739 Fig. 5 shows that the election of the Moving Average affects to the system positioning accuracy,
740 however, it is not crucial since the differences of the accuracy are in the range of 0.05 m, while the
741 difference between 0.5 and 0.95-percentiles is quite higher (around 0.1 m). Moreover, the results show
742 that the accuracy slightly depends on the sampling rate of on-line calibration measurements.
743 3.5. Performance evaluation and testing

744 The last experiments evaluate the system working in fully operative mode, i.e., the full
745 fingerprints map is updated in parallel to positioning operations. For on-line calibration
746 measurements, 8 devices are situated in the room, which send testing signals each 0.5 s (120
747 samples/min.) After measuring the RSS of the on-line calibration devices and taking into account the
748 off-line map, the system re-calculates the fingerprints map.
749 In these conditions, the system accuracy is calculated as follows: a moving testing device with
750 well-known position emits signals which are received and measured by the 6 transceivers, so the
751 testing device may be localized. The device is tracked in two ways: by using the fingerprints map
Sensors 2018, 18, x FOR PEER REVIEW 20 of 23

752 (which takes into account off-line and on-line calibration) and by using the off-line calibration map
753 only, so that we may compare the results and understand the effect of on-line calibration.
754 In the case that the difference between real position and estimated position is lower than 0.2√2
755 (the fingerprints map is a 0.4 m-grid map), then the test result is considered as positive. The accuracy
756 is the relation between positive results and all the test results. The test has been performed during
757 one working day and the test measurements have been grouped following one-hour intervals, as
758 presented in Fig. 6. Notice that each test value of the figure refers to the accuracy ratio for all the
759 experiments conducted in given hour.
Localisation accuracy (%) 100

80

60

40 Off-line calibration map

20 Full fingerprints map

0
10:00

11:00

12:00

13:00

14:00

15:00

16:00

17:00

18:00

19:00

20:00
9:00

One-hour intervals of measurements

760 Figure 6. Positioning accuracy during one-day normal functioning (full fingerprints map and off-line
761 only calibration map).

762 The results show the supremacy of the full fingerprints map (which takes into account
763 environmental changes) over the off-line calibration map. In some tests, the accuracy is increased up
764 to 32 %. The accuracy of the system when the full fingerprints map is used is very high (over 80% in
765 all the tests). The accuracy decreases in the hours when more clients were in the shop due to the
766 increase of obstacles and interferences. Logically, the differences among our system and off-line-
767 based systems are higher in hours of high influx of customers .

768 4. Conclusions and future work


769 This paper shows the deployment of an indoor positioning system that includes multi-
770 technology multi-antenna infrastructure and considers calibration and positioning processes running
771 in parallel for usage in general stores. The system fills the gap between static and dynamic solutions
772 thanks to the on-line calibration, which allows to update the calibration map to changing
773 environment conditions. In fact, the fingerprint solutions proposed until now showed good results
774 in stationary conditions however, they failed when uncertainties arise. The system has been deployed
775 and tested in a shop. The results show outstanding accuracy (in comparison to other existing
776 approaches) and capability of adapting to the ongoing changes in environment. A distinctive feature
777 of the proposed solution is the use of single transceiver modules (multi-antenna system) for several
778 technologies. This is possible thanks to the fact that the technologies referred in the paper operate in
779 the same frequency band, and in addition the proposed scheme will only be used for the needs of
780 users’ devices positioning (no other data requiring high bandwidth will be sent). Another innovative
781 feature is the separation between the functions responsible for broadcast of the radio channel and the
782 control functions for the radio transmission. Furthermore, the solution is based on virtualization and
783 introduce modularity, both these features ensure flexibility for future deployments. The result is a
784 system where many transceiver modules monitor the area so that the system’s cost is reduced.
785 Based on our system implementation, we conducted extensive testbed experiments directed to
786 analyze the effect of off-line and on-line calibration as well as to show the system accuracy.
787 All the experiments provided in this text aimed to demonstrate the validity of the solution and
788 to make comparison with other solutions feasible (e.g., the ones provided in [30]), which required
789 best conditions environment (i.e., one testing device). Our future work is directed to understand the
Sensors 2018, 18, x FOR PEER REVIEW 21 of 23

790 limits of the system (accuracy) when multiple devices based on different technologies are considered.
791 For this scope, we will provide different fingerprint maps for different technologies, such that the
792 system checks the technology of the customer’s device and tries to position it with the specific
793 technology fingerprints map. Moreover, our future work will be directed to develop the
794 communication framework between the positioning system and the on-line testing devices in order
795 that the system is capable of adjusting automatically the on-line calibration frequency.
796 Funding: This work was undertaken under the AAL vINCI project supported, among others, by the National
797 Centre for Research and Development in Poland and the Research Promotion Foundation in Cyprus.

798 Conflicts of Interest: The authors declare no conflict of interest.

799 Author contribution: Conceptualization, Jordi Mongay Batalla and Jose Angel Flores; Investigation, Jordi
800 Mongay Batalla, Constandinos X. Mavromoustakis and George Mastorakis; Methodology, Jordi Mongay Batalla;
801 Writing – original draft, Jordi Mongay Batalla.

802 References
803 1. I. Sabek, M. Youssef, A.V. Vasilakos, ACE: An Accurate and Efficient Multi-Entity Device-Free WLAN
804 Localization System. IEEE Trans. Mob. Comput. 14(2): pp 261-273, 2015
805 2. M.Z.A. Bhuiyan, G. Wang, A.V. Vasilakos, Local Area Prediction-Based Mobile Target Tracking in Wireless
806 Sensor Networks. IEEE Trans. Computers 64(7), pp 1968-1982, 2015
807 3. H. Choi, H. Jin, S. Chan Kim, “RSS Bias Compensation in BLE Beacon Based Positioning System”, Ninth
808 International Conference on Ubiquitous and Future Networks (ICUFN) , Milan Italy, July 2017.
809 4. S. Sahin, H. Ozcan, K. Kucuk, “Smarttag: An Indoor Positioning System Based on Smart Transmit Power
810 Scheme Using Active Tags”, IEEE access vol. 6, April 2018, pp. 23500 – 23510.
811 5. L. Khalil, A. Waadt, G. Bruck, P. Jung, “Positioning Framework for WLAN 802.11n Utilizing Kalman Filter
812 on Received Signal Strength”, International Wireless Communications and Mobile Computing Conference
813 (IWCMC), Nicosia, Cyprus, Aug. 2014.
814 6. J. Yim, Introducing a decision tree-based indoor positioning technique. Expert Systems with Applications,
815 34(2), 2008, pp. 1296-1302.
816 7. S. Rehman, K. W. Sowerby, S. Alam, I. Ardekani, “Radio Frequency Fingerprinting and its Challenges”,
817 IEEE Conference on Communications and Network Security, CA. USA, Oct. 2014.
818 8. D. V. Le, N. Meratnia, P. J.M. Havinga,” Unsupervised Deep Feature Learning to Reduce the Collection of
819 Fingerprints for Indoor Localization using Deep Belief Networks”, International Conference on Indoor
820 Positioning and Indoor Navigation (IPIN), Nantes France, Sep. 2018.
821 9. S. Tomažič, D. Dovžan, I. Škrjanc, “Confidence-interval Fuzzy Model-based Indoor Localization”, IEEE
822 Transactions on Industrial Electronics, June 2018, pp. 2015-2024.
823 10. K. Youssef, L.-S. Bouchard, K. Z. Haigh, J. Silovsky, B. Thapa, C. P. Vander Valk, “Machine Learning
824 Approach to RF Transmitter Identification”, IEEE Journal of Radio Frequency Identification, Nov. 2018.
825 11. J. Kim, D. Han, “Passive WiFi Fingerprinting Method”,International Conference on Indoor Positioning and
826 Indoor Navigation (IPIN), Nantes France, Sep. 2018.
827 12. P. Bahl and Y. Padmanabhan, "Radar: an in-building rf-based user location and tracking system," in INFO
828 COM 2000. Nineteenth Annual Joint Coriference of the IEEE Computer and Communications Societies, vol.
829 2, 2000, pp. 775-784.
830 13. A. Ettlinger, G. Retscher, “Positioning using ambient magnetic fields in combination with Wi-Fi and RFID”,
831 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Alcalá de Henares, Oct.
832 2016.
833 14. R. Xu, W. Chen, Y. Xu and S. Ji, A New Indoor Positioning System Architecture Using GPS Signals. Sensors
834 2015, 15(5)
835 15. Z. Farid, R. Nordin, and M. Ismail, Recent Advances in Wireless Indoor Localization Techniques and
836 System. Journal of Computer Networks and Communications, 2013
837 16. L. Luoh, ZigBee-based intelligent indoor positioning system soft computing, Soft Computing, vol. 18, Issue
838 3, pp 443–456, March 2014
839 17. Q. Yang, Z. Bian, “A Metric Learning Model for Localization”, IEEE Advanced Information Management,
840 Communicates, Electronic and Automation Control Conference (IMCEC), Xi'an, China, Oct. 2016.
Sensors 2018, 18, x FOR PEER REVIEW 22 of 23

841 18. X. Chen, Z. J. Wang, “Reliable Indoor Location Sensing Technique Using Active RID”, 2nd International
842 Conference on Industrial Mechatronics ad Automation, Vancouver Canada, 2010.
843 19. G. Murad Reis, H. León, T. Alam, J. Anderson, L. Bobadilla, R. N. Smith, “A Whitening-based Tracking
844 Algorithm for Autonomous Underwater Vehicles”, OCEANS - MTS/IEEE Kobe Techno-Oceans (OTO),
845 Kobe Japan, May 2018.
846 20. Y-K Kim, S-H Choi and J-M Lee, “Enhanced Outdoor Localization of Multi-GPS/INS Fusion System Using
847 Mahalanobis Distance”, International Conference on Ubiquitous Robots and Ambient Intelligence (URAI),
848 Jeju Korea, Nov 2013.
849 21. Y. Gu, A. Lo, I. Niemegeers, A survey of indoor positioning systems for wireless personal networks. IEEE
850 Communications Surveys & Tutorials, vol. 11, issue 1, First Quarter 2009
851 22. D. Ksentini, A.R. Elhadi, N. Lasla, Inertial Measurement Unit: Evaluation for Indoor Positioning.
852 International Conference on Advanced Networking Distributed Systems and Applications (INDS), 2014.
853 23. K. Al Nuaimi, H. Kamel, A Survey of Indoor Positioning Systems and Algorithms. International Conference
854 on Innovations in Information Technology (IIT), 2011
855 24. R. Faragher and R. Harle (2014) “An Analysis of the Accuracy of Bluetooth Low Energy for Indoor
856 Positioning Applications”. Proceedings of the 27th International Technical Meeting of The Satellite Division
857 of the Institute of Navigation (ION GNSS+), 2014, Florida, US.
858 25. A.K.M. Mahtab Hossain and Wee-Seng Soh. A comprehensive study of bluetooth signal parameters for
859 localization. In IEEE 18th International Symposium on Personal, Indoor and Mobile Radio
860 Communications, 2007. PIMRC 2007
861 26. M. Ji, J. Kim, J. Jeon, Y. Cho, Analysis of positioning accuracy corresponding to the number of BLE beacons
862 in indoor positioning system. 17th International Conference on Advanced Communication Technology
863 (ICACT), 2015
864 27. D. Čabarkapa, I. Grujić, P. Pavlović, Comparative analysis of the Bluetooth Low-Energy indoor positioning
865 systems. 12th International Conference on Telecommunication in Modern Satellite, Cable and Broadcasting
866 Services (TELSIKS), 2015
867 28. R. Dobbins, S.Garcia, B. Shaw Software Defined Radio Localization Using 802.11-style Communications A
868 Major Qualifying Project Report Submitted to the Faculty of Worcester Polytechnic Institute
869 29. T Vaupel, J. Seitz, F. Kiefer, S. Haimerl, and J. Thielecke, "Wi-fi positioning: System considerations and
870 device calibration," in 2010International Conference on Indoor Positioning and Indoor Navigation(IPIN),
871 2010.
872 30. D. Lymberopoulos, J. Liu, X. Yang, R.R. Choudhury, S. Sen, V. Handziski, Microsoft indoor localization
873 competition: Experiences and lessons learned. GetMobile, vol. 18, Issue 4. October 2014.
874 31. Nextome Co.: Indoor Positioning and Navigation System. https://www.nextome.net/ (last access: July
875 2017).
876 32. Z. Jiangy, W. Xiy, X.-Y. Li, J. Zhaoy, and J. Hany. HiLoc: A TDoA-Fingerprint Hybrid Indoor Localization
877 System. Technical report, Microsoft Indoor Localization Competition, 2014.
878 33. W. Burakowski et al., Provision of End-to-End QoS in Heterogeneous Multi-Domain Networks. Annals of
879 telecommunications - annales des télécommunications, Vol.: 63, Issue: 11, Pages: 559-577. 2008.
880 34. J. M. Batalla, G. Mastorakis, C.X. Mavromoustakis, J. Zurek “On cohabitating networking technologies with
881 common wireless access for home automation system purposes”. IEEE Wireless Communications Vol. 23,
882 Is. 5. October 2016
883 35. C. X. Mavromoustakis, J. Mongay Batalla, G. Mastorakis, E. Markakis, E. Pallis, Socially-oriented Edge
884 Computing for Energy-Awareness in IoT Architectures, IEEE Communications Magazine. July 2018.
885 36. J. Mongay Batalla and F. Gonciarz, Deployment of Smart Home management system at the edge:
886 mechanisms and protocols, Neural Computing and Applications. June 2018.
887 37. J. Mongay Batalla, A. Vasilakos and M. Gajewski, Secure Smart Homes: Opportunities and Challenges,
888 ACM Computing Surveys. October 2017.
889 38. J. Mongay Batalla, M. Gajewski, W. Latoszek, P. Krawiec, C. Mavromoustakis, G. Mastorakis, ID-based
890 service-oriented communications for unified access in IoT, Elsevier Computer & Electrical Engineering
891 Journal, vol.: 52, issue: 2016, Pages: 98-113, 2016.
892 39. N. LaSorte, S. Rajab, and H. Refai, “Experimental assessment of wireless coexistence for 802.15.4 in the
893 presence of 802.11g/n,” in IEEEEMC ’12, 2012, pp. 473–479.
Sensors 2018, 18, x FOR PEER REVIEW 23 of 23

894 40. C. Pendão, A. Moreira. "FastGraph - Organic 3D Graph for Unsupervised Location and Mapping," 2018
895 International Conference on Indoor Positioning and Indoor Navigation (IPIN), Nantes, 2018.
896 41. F. Potortì, A. Crivello, M. Girolami, E. Traficante and P. Barsocchi, "Wi-Fi probes as digital crumbs for
897 crowd localisation," 2016 International Conference on Indoor Positioning and Indoor Navigation (IPIN),
898 Alcala de Henares, 2016.

You might also like