LOCATION Based Service (LBS) has become an indispensable

Transcription

1 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 1 WinIPS: WiFi-based Non-intrusive Indoor Positioning Syste with Online Radio Map Construction and Adaptation Han Zou, Ming Jin, Hao Jiang, Lihua Xie, Fellow, IEEE and Costas J. Spanos, Fellow, IEEE Abstract WiFi fingerprinting-based Indoor Positioning Syste (IPS) has becoe the ost proising solution for indoor localization. However, there are two ajor drawbacks that haper its large-scale ipleentation. Firstly, an offline site survey process is required which is extreely tie-consuing and labor-intensive. Secondly, the RSS fingerprint database built offline is vulnerable to environental dynaics. To address these issues coprehensively, in this paper, we propose WinIPS, a WiFi-based non-intrusive IPS that enables autoatic online radio ap construction and adaptation, aiing for calibration-free indoor localization. WinIPS can capture data packets transitted in existing WiFi traffic and extract the RSS and MAC addresses of both WiFi Access Points (APs) and obile devices in a nonintrusive anner. APs can be used as online reference points for radio ap construction. A novel Gaussian process regression odel is proposed to approxiate the non-unifor RSS distribution of an indoor environent. Extensive experients were conducted, which deonstrated that WinIPS outperfors existing solutions in ters of both RSS estiation accuracy and localization accuracy. Index Ters Indoor Positioning Syste (IPS), radio ap construction and adaptation, WiFi, Gaussian process regression. I. INTRODUCTION LOCATION Based Service (LBS) has becoe an indispensable part of our daily lives due to its widespread applications, e.g., navigation, advertiseent, shopping, etc., in sart buildings. The quality of LBS largely depends on the localization accuracy [1]. Global Positioning Syste (GPS) can provide satisfactory localization accuracy for ost outdoor LBS. However, it is incapable of providing sufficient localization accuracy in indoor environents due to the lack of line of sight (LoS) propagation channel. Therefore, a lot of efforts have been devoted to developing Indoor Positioning Systes (IPSs) in the past two decades [1] [3]. Aong the This work is funded by the Republic of Singapore National Research Foundation (NRF) through a grant to the Berkeley Education Alliance for Research in Singapore (BEARS) for the Singapore-Berkeley Building Efficiency and Sustainability in the Tropics (SinBerBEST) Progra. BEARS has been established by the University of California, Berkeley as a center for intellectual excellence in research and education in Singapore. The work is also partially funded by Republic of Singapore NRF under grant NRF13EWT-EIRP- 1 and NRF11NRF-CRP1-. H. Zou, M. Jin, and C. J. Spanos are with the Departent of Electrical Engineering and Coputer Sciences, University of California, Berkeley, CA 7 USA (e-ail: hanzou, jining, spanos@berkeley.edu). L. Xie is with the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 37 (e-ail: elhxie@ntu.edu.sg). H. Jiang is with the College of Electrical Engineering and Autoation, Fuzhou University, Fuzhou, China (e-ail: jiangh@fzu.edu.cn). proposed techniques, WiFi has been acknowledged as the ost proising alternative to GPS for indoor localization because coercial off-the-shelf (COTS) WiFi devices and infrastructures are widely available in indoor environents and ost of obile devices (MDs) are equipped with WiFi odules. Fingerprinting-based localization algorith is the ost widely adopted algorith for WiFi-based IPS due to its ability to capture signal variances in coplex indoor environents ore accurately than other algoriths [] []. However, there are two ajor drawbacks that restrain the for large-scale ipleentation. One is that the offline site survey process is extreely tie-consuing and labor-intensive. Multiple RSS saples need to be easured at nuerous calibration points to ensure localization accuracy. The other is that the offline calibrated database is vulnerable to environental dynaics [7], as the real-tie RSS readings collected during the online localization phase can deviate fro those stored in the offline radio ap due to variation in teperature, huidity, occupancy distribution and ultipath effects. Serious localization errors ay be introduced if the radio ap is not updated adaptively. Previous works have tried to use an indoor radio propagation odel for online radio ap construction to replace the laborious offline site survey process [], []. However, the siple log-distance path loss odel fails to capture the non-unifor RSS distribution in coplex indoor environents. Soe works deploy fixed reference anchors to obtain real-tie RSS readings for radio ap adaptation [1], [11]. Nevertheless, the requireent of extra hardware ipleentation is the bottleneck of these ethods. Learningbased approaches are also introduced to reduce the nuber of reference anchors to be deployed [1], [13]. However, these ethods still need to conduct an offline initialization phase to collect RSS fingerprints as label data for learning purposes. Although certain crowdsourcing ethods have been introduced in [1], [1] recently to tackle the issues entioned, extra user intervention is required. Therefore, an efficient, easily ipleentable and non-intrusive schee for online radio ap construction and adaptation is urgently needed. In this paper, we propose, WinIPS, a WiFi-based nonintrusive indoor positioning syste that enables autoatic online radio ap construction and adaptation for calibrationfree indoor localization to overcoe the aforeentioned issues of WiFi fingerprinting-based IPS. For RSS data acquisition, we develop WinSMS, a novel intelligent wireless syste that can capture data packets transitted in existing WiFi traffic (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

2 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications and extract the RSS and MAC addresses of both APs and MDs in a non-intrusive anner without introducing any extra hardware. Since we can obtain the real-tie RSS easureents of APs, they becoe natural online reference points for online radio ap construction and adaptation. Therefore, we can copletely avoid the tedious offline site survey process. Furtherore, in order to build up a ore fine-grained radio ap, we propose the Gaussian Process Regression (GPR) with Polynoial Surface Fitting Mean (PSFM-GPR), a reliable regression technique dedicated to predict RSS on virtual reference points (VRPs). It can well capture non-unifor RSS distributions over coplex indoor environents. PSFM-GPR odels the RSS distribution with a two-diensional surface which is closer to practical scenarios. Moreover, this online radio ap better adapts and is robust to environental dynaics than the traditional offline calibrated RSS database since it is up-to-date all the tie. Since the online radio ap is based on AP generated RSS values, it is not suitable for localizing MDs directly due to the device heterogeneity issue. Instead of raw RSS values, we leverage Signal Tendency Index (STI) [1], which copares the shapes of RSS vectors between RSS readings of MD and online RSS fingerprint database. Then, we propose Signal Tendency Index - Weighted K Nearest Neighbor (STI-WKNN), that adopts the siilarity index STI as a novel weighting schee for WKNN, to iprove the localization accuracy of WinIPS across heterogeneous devices. Extensive experients were carried out over a duration of six onths to validate the effectiveness of WinIPS in a real-world ulti-functional office. The experiental results deonstrate that PSFM-GPR achieves a. db average RSS estiation accuracy and a 1.71 average localization accuracy, which outperfors the existing approaches, such as GPR with Log- Distance Mean (LDM-GPR) [] and Geography Weighted Regression (GWR) [17]. Furtherore, STI-WKNN iproves the localization accuracy by 3.% over traditional algoriths across heterogeneous MDs. In suary, we ake the following contributions: We develop a WiFi-based non-intrusive IPS, WinIPS, that is able to estiate locations of obile devices without app installation on the user s side. For online RSS data acquisition, we design WinSMS to overhear WiFi traffic and extract RSS values and MAC addresses of obile devices and APs fro the data packets in a non-intrusive anner. WinSMS can be directly ipleented on COTS WiFi routers, aking the natural reference points without introducing any extra hardware infrastructure. For online radio ap construction and adaptation, we propose PSFM-GPR, which is able to build up and update fine-grained radio ap autoatically over environental dynaics and discard the ipractical laborious offline site survey process. We introduce STI-WKNN that allows WinIPS to provide a high localization accuracy consistently across heterogeneous obile devices. We prototype WinIPS and test it in real coplex indoor environent. Proising results indicate that WinIPS akes substantial progress towards fortifying WiFi fingerprint-based IPS for feasible large-scale coercialization. The rest of the paper is organized as follows. The related work is briefly reviewed in Section II. Section III introduces the detailed syste design of WinIPS, as well as the ethodologies of WinSMS, PSFM-GPR and STI-WKNN. In Section IV, our experiental testbed and data collection procedure are described first, and experiental results and perforance evaluation of WinIPS are then reported. We conclude this paper with Section V. II. RELATED WORK In this section, we first present a brief overview on fingerprinting-based localization algoriths and their liitations, and then introduce existing approaches that try to tackle the probles. A. Liitations of Fingerprinting-based Localization Algoriths Fingerprinting-based localization algoriths can be classified into two categories: deterinistic approaches [], [11] and probabilistic approaches [1], [1]. Pioneered by RADAR [], deterinistic approaches easure the difference between realtie RSS saples and the ean of RSS fingerprints, calculating the ost atched fingerprints. They can provide eterlevel localization accuracy with a dense radio ap. On the other hand, probabilistic approaches calculate the likelihood between the real-tie RSS saples and RSS distributions of fingerprints stored in the database. Statistical techniques such as axiu likelihood estiation [1], axiu a posteriori estiation [1] and Gaussian process [] are eployed to estiate the user location. Several published results have shown that the fingerprintingbased localization algoriths outperfor other ethods, such as the tie-of-arrival, angle-of-arrival and odel-based approaches [1]. Soe detailed perforance analysis of RSS fingerprinting-based localization algoriths, such as Crar- Rao lower bound, are elaborated in [1], []. There are two ajor drawbacks of the existing fingerprinting-based algoriths. One is that the offline site survey process is tie consuing, labor exhaustive and expensive. In order to achieve sufficient localization accuracy, the WiFi RSS fingerprints fro different access points (APs) need to be easured at a huge nuber of calibration points, which is ipractical for large indoor environents such as shopping alls, stadius and airports. The other drawback is that the offline calibrated RSS fingerprint database is vulnerable to environental dynaics [7]. RSS is known to be susceptible to various environental changes including instant interference, such as the opening and closing of doors and oving etal objects, as well as continuous interference, such as variations in teperature, huidity and occupancy distribution. Another source of interference is ultipath effects, which include reflection, diffusion and diffraction in indoor environents. As a consequence, the real-tie RSS saples collected during the online localization phase can severely deviate fro those (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

3 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 3 stored in the offline radio ap, leading to serious location errors. In suary, the laborious and prolonged offline site survey process and the vulnerability to environental dynaics of fingerprinting-based approaches hinder its further coercialization. B. Radio Map Construction and Adaptation Several schees have been proposed to reduce the anual efforts for offline site survey and update the radio ap online, including fixed reference anchor ethods [1], [11], [17], [3], calibration-free ethods [], [], learning-based ethods [1], [13] and crowdsourcing ethods [1], [1]. Specifically, LANDMARC [1] and LEASE [11] developed an adaptive offset of the RSS variations by eploying reference anchors deployed at known fixed locations with real-tie RSS observations. Nonetheless, these approaches require a very dense deployent of reference anchors to construct the radio ap accurately. In [17], self-ade WiFi anchors are introduced to obtain real-tie RSS observations and Geography Weighted Regression (GWR) is adopted for online radio ap construction to reduce the workload for offline site survey. It is noted that all these ethods still require extra hardware to be deployed and are infeasible for largescale ipleentation. A calibration-free ethod which uses an indoor radio propagation odel for online radio ap construction to reove the offline site survey process is presented in []. Nevertheless, the siple log-distance path loss odel cannot describe the coplex RSS distribution precisely. In [], the idea of eploying RSS data aong APs to establish a radio ap and using the GPR with Log-Distance path loss odel for RSS odeling is introduced. However, they fail to odify the AP firware due to its technical difficulty and instead put wireless onitors beside each AP. As a result, an extra device is still needed. Several learning-based approaches are also introduced to reduce the nuber of reference anchors to be deployed [1], [13], []. LEMT [1] perfored radio ap adaptation by training the functional relationship between each location and its neighboring locations based on nonlinear regression analysis and the odel tree ethod, since neighboring locations have highly correlated RSS characteristics in general. The drawback of LEMT is that the process of building huge nubers of trees in each RSS sniffing period is tieconsuing, which akes it difficult for real-tie applications. Other learning techniques such as ulti-view learning [] and anifold alignent [13] are also utilized to transfer RSS inforation across different ties and devices. Nevertheless, they still need to collect certain nubers of offline RSS fingerprints as label data for learning purpose. Crowdsourcing ethods, which eploy the full sensing capabilities of MDs, are introduced to reduce the efforts for radio ap construction as well [1], [1]. Zee [1] utilized inertial easureent unit (IMU), coprised of acceleroeters, gyroscopes and agnetoeters, and RSS reading fro the MDs to build up a radio ap. Walkie-Markie [1] used landarks, such as turns, escalators and elevators, to enhance crowdsourcing perforance. Nevertheless, extra user intervention is needed for these approaches and continuous IMU Fig. 1. WinIPS syste architecture, illustrating odules of RSS data acquisition, online radio ap construction and localization. onitoring will consue a lot of MDs batteries, which is an ipractical solution. A. Syste Overview III. SYSTEM DESIGN The objective of WinIPS is to realize autoatic online radio ap construction and adaptation for calibration-free indoor localization. The syste architecture of WinIPS is illustrated in Fig. 1. It consists of three ain parts: RSS data acquisition, online radio ap construction and online localization. For RSS data acquisition, we develop the WiFi-based non-intrusive Sensing and Monitoring Syste (WinSMS), which enables COTS WiFi APs to intercept the data packets transitted in the existing WiFi traffic and extract RSS values in a non-intrusive anner without extra hardware infrastructure. All the data will be forwarded to a back-end server for radio ap construction and localization. We propose PSFM-GPR, a reliable regression technique dedicated for RSS predictions on each VRP to construct and update a fine-grained online RSS radio ap over various environental dynaics. For online localization, STI- WKNN is adopted to estiate the locations of heterogeneous MDs with consistent high localization accuracy. The users can use any browser on their MDs to obtain the estiated location through the WinIPS Web server without the need of installing an app. The following sections will introduce the ethodologies of WinSMS, PSFM-GPR and STI-WKNN, respectively. B. WinSMS for RSS Data Acquisition The ain drawbacks of fingerprinting-based approaches, the laborious offline site survey process and the vulnerability to environental dynaics have been elaborated in Section II. In addition, Apple Inc. has not provided any RSS API for thirdparty developers. Due to these reasons, active WiFi scanning via MD is not a practical ethod for establishing radio aps anyore. Therefore, it is urgent and indispensable to design a schee for online RSS radio ap construction and adaptation in an accurate, reliable, efficient, practical and non-intrusive anner (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

4 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications TABLE I ONLINE RSS OBSERVATIONS AMONG APS CAPTURED BY WINSMS. AP 1 AP AP n AP 1 RSS AP(1,1) RSS AP(1,) RSS AP(1,n) AP RSS AP(,1) RSS AP(,) RSS AP(,n)... AP n RSS AP(n,1) RSS AP(n,) RSS AP(n,n).. -3 AP Fig.. WinSMS syste to collect RSS data fro counication aong APs and obile devices. To overcoe this bottleneck, we develop, WinSMS, an intelligent wireless syste that enables COTS WiFi APs to overhear the data packets transitted in the existing WiFi traffic in real-tie without any intrusion on the user side. It can be ipleented on ost of the COTS WiFi routers that support the OpenWrt [] operating syste. WinSMS can create a WiFi LAN to provide basic Internet services for users in its wireless network coverage. More iportantly, it has the ability to overhear the data packets transitted between each MD and WiFi routers, and accurately retrieve the RSS values and corresponding MAC addresses as identifiers. Then, all the inforation will be sent to a back-end server without requiring user to install any dedicated app for data acquisition. Fig. presents the syste architecture of WinSMS. The ain coponents of WinSMS includes the COTS WiFi APs, a back-end server, as well as users and their MDs. All the APs in WinSMS perfor the following ajor tasks: capture the.11n data packets in the network, extract relevant inforation fro the packets, arrange the in a particular forat and forward the to the back-end server. We upgrade the firware of APs with OpenWrt and add a designed software based on Libpcap [] to sniff existing WiFi traffic, and capture as well as analyze the data packets. Unlike traditional active RSS scanning via a MD which has a liited sapling rate, APs are able to overhear sustainable aount of data packets generated by various existing applications on MDs, such as data strea fro watching videos, push notification services and periodic eail fetching, at the axiu rate around 1 packets per second in a non-intrusive anner. Furtherore, since WinSMS opportunistically captures the data packets fro existing WiFi traffic, it poses no additional burden on the battery life of MD. Noticing that usually a person cannot ove a significant distance in a second and the RSS value cannot change draatically in such a short tie, the RSS values received within 1 second are averaged out as a pre-filtering step. In this way, the RSS values collected by WinSMS are soother than those by the active scanning ethod. The weakest signal strength is set to be db. If a particular data packet is received by only one AP, we set the value received by the others as db which effectively eans that the device is outside the range of that AP. After that, the retrieved RSS values of MDs with their corresponding AP AP3 AP AP AP AP AP AP1 AP AP3 AP AP AP AP7 AP Fig. 3. Visualization of pairwise RSS atrix aong APs (db). For instance, the RSS easureent between AP and AP is -7 db. As discussed in the ain text, we assue the self-sensed RSS of each AP is -3 db. MAC addresses will be sent to the back-end server through the UDP protocol. The server is responsible for parsing the data and building up the online RSS fingerprint database for localization. For each AP, in addition to capturing the data packets sent and received by each MD, it can overhear packets of other APs as well. Therefore, the RSS easureents at these APs can be leveraged for online ap construction. As suarized in Table I, all the APs can be used as natural online reference points for RSS radio ap construction and adaptation since we have their physical coordinates and real-tie RSS readings. Fig. 3 deonstrates the visualized pairwise RSS of APs. In principle, each AP cannot sense the signal strength of itself. Therefore, we calibrate the average RSS of two APs placed side-by-side and assign 3 db as the self-sensed RSS to coplete the pairwise RSS atrix of APs as shown in Fig. 3. As shown in Fig., the RSS values on the liited nubers of APs ay not be good enough to describe a fine-grained RSS distribution for each AP. In order to obtain a ore fine-grained radio ap, we introduce VRPs and propose the PSFM-GPR, a suitable RSS odeling schee that is able to accurately estiate the RSS values of each AP at predefined VRPs for fine-grained online radio ap construction. The ethodology of the PSFM-GPR is introduced in the following section (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

5 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications for each AP consist of pairs of (l i, s i ) n i=1, l i L, s i S, where l i = (x i, y i ) is the two-diensional coordinates of an AP, and s i is an RSS value of the AP at location l i. The relationship between the two-diensional space L and RSS S can be odeled as a GP: s i = f(l i ) + ɛ i where ɛ i is independent and identically distributed (i.i.d.) additive zero-ean Gaussian noise with variance σɛ. Assue the RSS observations at each AP can be drawn fro the GP: s GP ( (l), k(l, l ) ) Fig.. Scenario of WinIPS ipleentation in a coplex indoor environent. Following the algorith outlined in Fig. 1, a fine-grained radio ap at virtual reference points is updated with WinSMS data to localize users in real-tie. C. PSFM-GPR for Online Radio Map Construction and Adaptation 1) Gaussian Process Regression Model for RSS Modeling: Adittedly, the RSS transitted fro a WiFi AP in a free space is a log linear delay function of the distance. Nevertheless, this property does not hold in practice due to the ultipath effects caused by furniture, walls and oving occupants in coplex indoor environents. Therefore, the ideal log-distance path loss odel is not able to predict the RSS distribution precisely anyore. An efficient and powerful nonlinear approach is required to odel the anoalous distribution of RSS values. As a nonparaetric nonlinear regression approach, GPR is an appropriate ethod for capturing the noisy nature of RSS, and predicting RSS values for online dynaic radio ap construction and adaptation. In fact, GPR has been widely eployed in nuerous areas, including geostatistics, spatial soothing, robotic applications and achine learning for probabilistic odeling, inference and prediction [7]. Moreover, previous works [], [] eployed GPR for RSS interpolation to reduce the nuber of reference points during the offline calibration phase. A Gaussian Process (GP) generates data located at any point of a finite set of rando variables Z which follows a joint ultivariate Gaussian distribution. It is characterized by its ean function (z) = E[f(z)] and the covariance function k(z, z ) = E[(f(z) (z))(f(z ) (z ))], where z Z. The arginalization property of GP [7] allows us to predict the posterior probability with an unknown input z according to soe given inputs z and their corresponding observations. Since the online radio ap construction process of each AP is siilar, we will explain how to use GPR to predict RSS values of AP i as an exaple, where AP i is one of the n APs. WinSMS enables each AP to scan not only the RSS of MDs, but also the RSS of other APs. Therefore, all the APs are natural online reference points (training points) for radio ap construction and adaptation. The corresponding dataset where ( ) and k(, ) represent the ean and covariance function of GP respectively. GP learns the covariance of the training dataset through the kernel covariance function. In our case, the input data are the two-diensional coordinates. The value of the kernel covariance function is higher when two points are near to each other and lower when two points are far away. We utilize the ost popular squared exponential kernel covariance function: [ l l k(l, l ) = σf ] exp r + σɛ δ(l, l ), (1) where σf and r are the hyperparaeters of GP and δ(, ) denotes the Kronecker delta function. Since we have n APs in the space, we can calculate the covariance of each pair of APs according to Equation (1) and obtain the n n covariance atrix K(L, L) for all pairs of training data. Suppose that we would like to predict RSS values {s j } j=1 S of AP i at VRPs {l j } j=1 L to build up a fine-grained radio ap. The ultivariate Gaussian distribution of training data and predicted RSSs with a zero-ean distribution can be described as follows: [ ] ( [ S K(L, L) + σ S N, ɛ I K(L, L ) K(L, L) K(L, L ) ]), where K(L, L ) is an n covariance atrix between S and S, and I is the identical atrix. The RSS value of this AP at an interested point l j can be predicted according to the posterior ean and variance of GP: s j = K(l j, L)[K(L, L) + σ ɛ I] 1 S, () cov(s j ) = K(l j, l j ) K(l j, L)[K(L, L) + σ ni] 1 K(L, l j ), where s j is the estiated ean RSS at this location, cov(s j ) denotes the posterior variance as an estiation confidence indicator, and K(l j, L) and K(L, l j ) are 1 n and n 1 atrices of the covariance between this point and all training points. As shown in Equation (), the GP odel usually adopts the zero-ean function (ZeroM-GPR) as the default settings, which eans that the estiated RSS values will tend to zero at locations that are far fro any training points (APs). This is obviously ipractical for RSS odeling. Previous works [] used the Log-Distance path loss odel to obtain a general ean of RSS and then ade use of GPR to estiate the (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

6 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications db AP Location Fig.. Illustration of RSS distribution of an AP in a coplex indoor environent, where the AP location is arked. residual RSS errors. The estiated RSS at an arbitrary location l j is calculated by s j = (l j ) + K(l j, L)[K(L, L) + σ ɛ I] 1 (S (L)), (l j ) = P L + 1α log( l j l APi /d ) + X, (3) where l j l APi indicates the distance between AP i and location l j, P L is the path loss coefficient of the RSS value at initial distance d, and α is the path loss gradient, and X represents the lognoral shadow fading with zero ean noise with standard deviation σ X []. These three paraeters of the Log-Distance path loss odel in Equation (3) are calculated by curve fitting with the training points. The Log-Distance Mean GPR (LDM-GPR) can be used to estiate the RSS distribution in an open space because it describes the relationship between RSS and distance. However, in practice, as shown in Fig., the RSS distribution is uch ore coplicated. The RSS values at the sae distance fro the AP are usually distinct due to ulti path effect and shadow fading results fro the obstacles that attenuate signal power through absorption, reection, scattering, and diffraction in coplex indoor environent. Hence, the LDM-GPR is no longer suitable since it does not consider the orientation or the surrounding environental property on each VRP. ) Online Radio Map Construction with PSFM-GPR: In order to address this issue, we propose the PSFM-GPR, which utilizes a two-diensional polynoial surface fitting odel to estiate the general ean of the RSS, and then utilizes GP to estiate the residual RSS errors. First of all, we assue the RSS distribution of AP i to be a two-diensional polynoial function as follows: (l) = β + β 1 x + β y + β 3 x + β y + β xy () where l = (x, y) denotes the coordinates of other APs. Since the WinSMS can obtain the RSS values of AP i at all other APs locations, all the paraeters β, β 1, β, β 3, β, β in Equation () can be estiated and updated online using two degree polynoial surface fitting. According to our data analysis regarding the fitting accuracy and the coputational. overhead, we found that two-degree polynoial surface fitting is good enough to capture the non-unifor RSS distribution. With this proper ean of RSS, the predicted RSS by the PSFM-GPR at any arbitrary location l j is calculated by s j = (l j ) + K(l j, L)[K(L, L) + σ ɛ I] 1 (S (L)), () (l j ) = β + β 1 x j + β y j + β 3 x j + β y j + β x j y j, () where (x j, y j ) are the coordinates of location l j. After estiating the RSS values of all the n APs by PSFM-GPR at the VRPs, we can obtain a RSS vector s j = [s 1 j, s j,..., sn j ], where 1 j and sj i (1 i n) denotes the RSS values fro AP i at each VRP l j. Therefore, a n RSS fingerprint database can be effectively built up online to avoid the cubersoe offline site survey process. 3) Online Radio Map Adaptation with PSFM-GPR: The radio ap adaptation is another crucial process of WinIPS syste because it keeps the radio ap up-to-date autoatically over various contextual dynaics including tie and space. Since the WinIPS syste can obtain the RSS values of all APs in real tie as presented in Table I, each colun in n n RSS atrix can be used as a trigger to deterine whether the syste should initiate the radio ap adaptation process for each AP. Algorith 1 Online radio ap adaptation algorith Initialization: Input: n - The total nuber of APs - The total nuber of VRPs s t 1 - n n RSS atrix of AP as shown in Table 1 s t 1 i - The RSS vector of AP i stored in the fingerprint database s t i - The RSS vector of AP i at the tie t θ th - The RSS threshold for AP RSSI differences Output: s t f - n Up-to-date RSS fingerprint database at tie t Check RSS profile of each AP: for i = 1,, n do if s t i st 1 i > θ th then RSS profile of AP i is required to update AP i AP Q else RSS profile of AP i is up-to-date end if end for Update RSS fingerprint database: for q = 1,, Q do AP q P SF M GP R to predict RSS on all VRPs for j = 1,, do s j = s j AP q end for end for return s t f The detailed procedure of radio ap adaptation is presented in Algorith 1. First of all, we will copare the differences (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

7 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 7 RSS (db) Location 1 Router iphone Galaxy S Nexus ipad Air Mi AP Index siilarities of the RSS curve shapes by using the ordinary Procrustes Analysis (PA) ethod [3] instead of using the raw RSS for fingerprint atching. To be specific, given a real-tie RSS vector fro a MD, s d, the translation step of the ordinary PA ethod will produce RSS (db) Location Router iphone Galaxy S Nexus ipad Air Mi AP Index Fig.. Deonstration of RSS for a router and obile devices, located at Location 1 (top plot) and Location (botto plot), easured by 1 APs (indexed in x-axis). The shapes of router and obile device RSS curves display siilar patterns. between the real-tie RSS values s t i and the RSS profile st 1 i stored in the database for all the APs. If the RSS distance between these two RSS vectors s t i st 1 i is larger than a RSS threshold θ th, it iplies that the RSS profile of AP i is outdated due to soe indoor environental dynaics, and the radio ap update procedure will be initiated for this AP. According to our epirical study, we set the threshold θ th to 1dB. The RSS values fro this AP at each VRP will be updated by the PSFM-GPR schee as introduced in Section III-C. In this way, the n online RSS fingerprint database will be up-to-date and be ore robust to various contextual dynaics copared to traditional offline fingerprint database. where s 1 d s d, s d s d,..., s n d s d (7) s d = 1 n n s i d. i=1 Then, in the unifor scaling step, we have where ŝ d = [s 1 d s d, s d s d,..., s n d s d ]/ˆσ, () ˆσ = 1 n n (s i d s d). i=1 The ŝ d is the transfored object of ordinary PA ethod. Siilarly, all the AP-based RSS vectors stored in the fingerprint database will be transfored as well. All the transfored RSS fingerprints {ŝ j } j=1 will be copared with ŝ d in ters of their shape siilarity. We define the Procrustes distance between the two vectors ŝ d and ŝ j, tered signal tendency index (STI), which is coputed by ST I j = ŝ d ŝ j () D. STI-WKNN Localization Algorith for Heterogeneous Mobile Device As introduced in the aforeentioned sections, an up-to-date and fine-grained online RSS fingerprint database is obtained using WinSMS and PSFM-GPR. However, the RSS values stored in this database are collected at APs. This database cannot be applied directly for localization of MDs because the RSS signatures of AP and MDs are usually different due to various heterogeneous factors such as distinct WiFi chipsets, WiFi antennas, hardware driver, and even operating systes [3], [31]. To illustrate this issue, we conducted an experient that collected RSS saples fro a TP-Link TL-WR73N portable router, as well as five different MDs: iphone, Galaxy S, Nexus, ipad Air and Mi at two identical locations with respect to 1 coodity WiFi APs in a coplex indoor environent. As observed in Fig., each curve connects the average RSSs between one device and 1 APs. The RSSs associated with router and MDs are significantly different, which verifies the effect of device heterogeneity. Therefore, the localization accuracy will be severely jeopardized if we eploy the RSS fingerprint database of a router (AP) to estiate the location of an MD directly. Meanwhile, another noteworthy observation fro Fig. is that the shapes of the curves display certain siilarities. In other words, one curve can be roughly recovered fro another one via translation and scale operations. Thus, to accoodate the device heterogeneity issue, we leverage the Signal Tendency Index (STI) [1], which copares the where denotes the Euclidean nor. After that, we introduce a new weighting schee which involves STI and integrate it with the classical localization algorith, Weighted K Nearest Neighbor (WKNN), naely STI-WKNN, instead of using the distance of RSS vectors as the weights. Since we have calculated the STI value ST I j between s d and each s j, a saller ST I j indicates that s j is siilar to s d. We further define a weight value w j for each s j, which is calculated as follows: w j = 1 ST I j j=1 1 (1) ST I j Then, the VRPs are sorted according to their w j in a descending order. Only top K VRPs and their corresponding physical coordinates are adopted to estiate the location of MD (x d, y d ), which is calculated by: (x d, y d ) = 1 c K (x k, y k ) w k (11) k=1 where (x k, y k ) denotes the coordinates of ith VRP and c = K k=1 wk is the noralization constant. In suary, the STI-WKNN localization schee first copares the siilarities of the RSS curve shapes between realtie RSS vector of a MD and those stored in the fingerprint database by the ordinary PA ethod. Then, the siilarity index STI is adopted as a novel weighting schee for WKNN to estiate the location of heterogeneous MDs (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

8 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications RSS variations of an AP in six onths RSS variations [db] RSS variations [db] RSS variations of an AP in one week (a) Layout of the testbed at T1. Fig.. Coparison of RSS variations (z-axis) in short-ter (left) and longter (right), showing uch larger long-ter variation. (b) Layout of the testbed at T. Fig. 7. Layout of the testbed (a) at the beginning of the experient (T1 ), and (b) six onths later after renovation (T ), showing significant indoor structure changes (shaded region). IV. E XPERIMENTAL R ESULTS AND D ISCUSSIONS A. Experiental Setup To validate the perforance of WinIPS coprehensively, extensive experients were conducted in a. 1. ulti-functional lab for six onths. The layout of the testbed at the beginning of the experient (labeled as T1 ) is depicted in Fig. 7(a), while the layout after renovation six onths later (labeled as T ) is presented in Fig. 7(b). As shown in Fig. 7, there are several obvious layout differences during the experient which definitely affect the RSS distribution in the area. We leveraged these changes to verify the radio ap adaptation and localization perforance of WinIPS under environental dynaics. This is different fro the traditional evaluation ethods [], [17] which usually adopt corridors or open spaces as testbeds, that are favorable for distance-related RSS odeling. As deonstrated in Fig. 7, our testbed includes workspaces, cubical offices, an open space for Unanned Aerial Vehicle (UAV) testing and a discussion roo. This coplex indoor environent is uch ore suitable than an ideal environent for perforance evaluation of WinIPS. In our experients, 1 TP-LINK TL-WR73N router, were adopted as APs for WinSMS in our experient. TLWR73N has a MHz Atheros AR7 CPU with MB flash eory and 3 MB RAM. The Atheros AR331 chipset is used in its platfor working on. GHz. To ipleent WinSMS, we upgraded the firware to OpenWrt and added our designed software. As shown in Fig. 7, the TL-WR73N nano router is sall in size and extreely easy to be deployed. We chose this router to deonstrate that coercial routers Fig.. Coparison of RSS estiation errors for different APs by ZeroMGPR [], LDM-GPR [], GWR [17] and PSFM-GPR. are becoing portable and easier for installation nowadays. Moreover, with the booing developent of Internet of Things (IoT), billions of IoT devices will be densely deployed in indoor environents for various purposes in the near future. Equipped with WiFi odules, they can be easily upgraded to serve as online reference points for dynaic radio ap construction and adaptation. The locations of these 1 APs are depicted in Fig. 7 and they were fixed on 1.-eter-high tripods to keep the on the sae height level. One server is eployed to process the RSS data sent by APs, construct and update the RSS radio ap and fingerprint database by PSFM-GPR, and adopts STI-WKNN to estiate the location of each MD. testing points (sall red circles in Fig. 7) were randoly selected to evaluate the perforance of WinIPS. To validate the RSS estiation accuracy of PSFM-GPR, we collected the real RSS values of a TL-WR73N router at these points as the ground truth. Furtherore, we also collected the RSS easureents of five MDs, including iphone, Galaxy S, Nexus, ipad Air and Mi, at all the testing points to evaluate the localization accuracy of STI-WKNN across heterogeneous devices. B. RSS Estiation Accuracy Firstly, we conducted an experient to continuously onitor the distribution of RSS variations of an AP (AP1) to understand the fluctuations of RSS caused by various environental dynaics in six onths. Fig. deonstrates the distribution of RSS variation of the AP over one week and six onths. As shown in Fig., the RSS variation over six (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

9 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications TABLE II C OMPARISON OF RSS ESTIMATION ERRORS FOR DIFFERENT METHODS, SHOWING THE MEAN AP AP AP AP AP AP AP AP THE STANDARD DEVIATION AP (a) ZeroM-GPR.7. σrss (σrss ) ( D B M ).. RSS estiation erros [db] 3 RSS estiation erros [db] 3 e RSS AP RSS estiation erros [db] RSS estiation erros [db] Method ZeroM-GPR LDM-GPR GWR PSFM-GPR (e RSS ) AND (b) LDM-GPR..7 (c) GWR (d) PSFM-GPR Fig. 1. Coparison of RSS estiation errors for AP by different ethods, showing the spatial distribution. For WinSMS, the real-tie RSS easureents aong the 1 APs collected by it can be suarized as a 1 1 RSS atrix, which is siilar to Table I. By eploying these data, we predicted RSS values fro APs by using PSFM-GPR at the testing points and copared it with the observed RSS (ground truth). Fig. and Table II copared the RSS estiation of PSFM-GPR with ZeroM-GPR, LDM-GPR [] and GWR [17] in ters of ean (e RSS ) and standard deviation (σrss ) of the RSS estiation error. The average estiated RSS error of PSFM-GPR is. db which is the sallest aong the four ethods. It is able to reduce the average RSS error by 7.%,.%, and.% copared to ZeroMGPR, LDM-GPR and GWR respectively. Moreover, the standard deviation of RSS error of PSFM-GPR is also the sallest aong the four ethods, indicating that RSS predicted by PSFM-GPR are ore stable than existing approaches. Furtherore, we evaluated the RSS estiation accuracy of PSFM-GPR in two-diensional space. To illustrate, Fig. 1 describes the estiated RSS error distribution of AP fro different RSS odeling ethods. As illustrated in Fig. 1(d), ost of RSS errors of PSFM-GPR are saller than 1 db and are distributed evenly in a low RSS error level. The reason for such an outstanding perforance is that PSFM-GPR perfors two-diensional surface fitting for RSS predictions, which well captures non-unifor RSS distributions in a different orientation. In contrast, the RSS errors of ZeroM-GPR is highest especially at the locations far away fro any AP (online reference points). On the other hand, LDM-GPR and GWR, failed to capture the non-unifor RSS distribution in coplex indoor environents because only the relationship 1 ZeroM-GPR LDM-GPR GWR PSFM-GPR.. CDF onths (long-ter) is uch larger than over one week (shortter). Thus, it indicates that the static radio ap calibrated at a particular tie is definitely unable to serve as the reference for consistent location estiation at all ties, since the real tie RSS values can vary significantly, especially in longter deployents. Radio ap adaptation strategies such as WinSMS is urgently desired to ake the IPS resilient to environental dynaics Localization error () Fig. 11. Cuulative distributions of localization error between different ethods. between RSS and distance in RSS odeling is considered. C. Localization Estiation Accuracy The aforeentioned section illustrates the RSS estiation evaluation of WinIPS. We present the localization accuracy evaluation of WinIPS in this section. To prepare a finegrained online RSS fingerprint database, the back-end server virtually divided our testbed into a grid and adopted the PSFM-GPR to predict RSS values fro all APs at the grid points (VRPs). The grid spacing between two adjacent VRPs was chosen to be around 1. according to the analysis in [33]. For evaluation, we collected RSS saples of each MD at each testing point, and used the average location estiated by STI-WKNN to copare with the physical location of each testing point (ground truth). 1) Coparison of Localization Accuracy Between Different Online RSS Prediction Methods: First of all, we evaluate the ipacts of different online RSS prediction ethods on localization accuracy. In the back-end server, we established three (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

10 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 1 Localization error distribution () of LDM-GPR (a) ZeroM-GPR Localization error distribution () of PSFM-GPR Localization error distribution () of GWR 1 1 Localization error distribution () of ZeroM-GPR (c) GWR. (b) LDM-GPR.7 (d) PSFM-GPR Fig. 1. Spatial distributions of localization errors (z-axis) by different ethods. Method ZeroM-GPR LDM-GPR GWR PSFM-GPR e LA () Iprove (%)...31 σla () Iprove (%) online RSS fingerprint databases using ZeroM-GPR, LDMGPR and GWR siilarly to that of PSFM-GPR. STI-WKNN was used as the localization algorith for all the schees in this evaluation to ake a fair coparison. The statistical attributes (i.e., the ean (e LA ) and standard deviation (σla ) of localization accuracy) via PSFM-GPR is copared with three other existing approaches. The overall perforance is suarized in Table III and Fig. 11. It is evident fro Table III that the localization accuracy of WinIPS is uch higher when PSFM-GPR is adopted for RSS prediction on VRPs. Fig. 1 depicts the distance error distribution in D over the floor plan of the four approaches. Siilar to the results shown in Table III, PSFM-GPR has the best perforance aong the four approaches. PSFM-GPR + STI-WKNN can provide a 1.71 average localization accuracy with the sallest σla =.3. It enhances the precision of indoor positioning by.% over ZeroM-GPR, 33.1% over LDM-GPR and.3% over GWR respectively. Furtherore, the sallest σla indicates that the online RSS fingerprint database generated by the PSFM-GPR can provide ore useful inforation for reliable localization service than the other approaches. We also explored potential correlations between the RSS estiation accuracy and the localization accuracy using PSFMGPR. Fig. 13 and Table IV copares the RSS estiation accuracy in ters of ean (e RSS ) and the standard deviation (σrss ) and the localization accuracy in ters of ean (e LA ) and standard deviation (σla ) when different nuber of APs are utilized. According to the analysis presented in [], the RSS variation is proportional to the square of the distance between routers (router density). It can be seen fro Fig. 13 that, RSS estiation errors becoe saller when ore APs are leveraged. Due to ulti path effect and shadow fading results fro the obstacles that attenuate signal power through absorption, reection, scattering, and diffraction in coplex indoor environent, the epirical results as shown in Fig. 13 and Table IVay not perfectly atch with the theoretical RSS Estiation error (db) TABLE III C OMPARISON OF LOCALIZATION ERRORS FOR DIFFERENT METHODS IN TERMS OF MEAN (e LA ) AND STANDARD DEVIATION (σla ). T HE IMPROVEMENT PERCENTAGE IS BASE - LINED AGAINST Z ERO M-GPR PERFORMANCE. 1 APs APs APs APs Linear fitting Localization error () Fig. 13. Plots of the RSS estiation error vs. the localization error, color coded by the nuber of APs in use, showing higher RSS estiation accuracy leads to ore accurate localization. TABLE IV C OMPARISON OF THE RSS ESTIMATION ACCURACY IN TERMS OF MEAN (e RSS ) AND THE STANDARD DEVIATION (σrss ) AND THE LOCALIZATION ACCURACY IN TERMS OF MEAN (e LA ) AND STANDARD DEVIATION (σla ) USING DIFFERENT NUMBER OF AP S. No. of AP 1 e RSS (db) σrss (db) e LA () σla () analysis as presented in []. However, as the result of the linear fitting, the general traces of the epirical results are siilar to the theoretical results, which validates that higher router density lead to saller variances of localization error. Thus, we conclude that there is a positive correlation between RSS estiation accuracy and localization accuracy for the PSFM-GPR. Another noteworthy point is that the results in Table III are coparable to those reported in [] which rely on a cubersoe offline calibrated RSS fingerprint database. ) Coparison between Traditional Offline Site Survey and PSFM-GPR: In this section, we copare the localization perforance of the PSFM-GPR to the traditional offline site survey ethod. We collected real RSS easureents of the MDs on the physical coordinates of each VRP to build up the offline (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

11 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 11 TABLE V COMPARISON OF LOCALIZATION ERROR IN TERMS OF MEAN (ē LA ) AND STANDARD DEVIATION (σ LA ) BETWEEN OFFLINE SITE SURVEY AND PSFM-GPR. Method ē LA () σ LA () Offline Site Survey (T 1 ) Offline Site Survey (T ) PSFM-GPR (T ) RSS fingerprint database. Although WinSMS is able to collect RSS values at a fast speed (. seconds/saple), we still spent hours to coplete the offline site survey process which is truly tie-consuing and labor-intensive. We perfored two offline site surveys and constructed the corresponding fingerprint databases at the beginning of the experient (T 1 ) and six onths later (T ). The testing data was collected on T. The overall perforance is presented in Table V. When the up-to-date online RSS fingerprint database generated by PSFM-GPR is copared with the offline database constructed on the sae day (T ), the average localization accuracy of it is only a little worse by.7% than the offline calibrated RSS fingerprint database. However, it is ipractical to build up an offline radio ap every day for localization purposes. To illustrate the vulnerability of the transitional offline site survey ethod to environental dynaics, we copared the perforance of an out-of-date offline RSS fingerprint database (T 1 ) to PSFM-GPR. Under this situation, PSFM-GPR reduces the localization error by.17% copared to the outdated offline RSS fingerprint database. In suary, PSFM-GPR can construct and update the RSS fingerprint database autoatically that enables WinIPS to provide consistent high localization accuracy over various environental dynaics. Furtherore, it avoids the cubersoe offline site survey process which is the ajor bottleneck for the large-scale coercialization of WiFi-based IPS. 3) Ipact of Device Heterogeneity: To validate the effectiveness of WinIPS under the ipact of device heterogeneity, we collected RSS easureents of five MDs at testing points for this evaluation. The overall results are suarized in Table VI. As observed fro Table VI, WinIPS can provide a high localization accuracy (within on average) consistently across heterogeneous MDs using STI-WKNN. Although the online RSS fingerprint database established by the PSFM- GPR is based on data aong APs, the device heterogeneity issue can be largely alleviated by coparing the siilarities of RSS curve shapes (STI) rather than the raw RSS values for WKNN fingerprint atching. Fig. 1 depicts the distance error distribution of the original WKNN and STI-WKNN. STI- WKNN has a uch better perforance in ters of localization accuracy copared to the original WKNN. It iproves localization accuracy by 3.% over the original WKNN across heterogeneous MDs. In suary, the erit of STI-WKNN enhances the robustness of WinIPS to device heterogeneity issues for indoor localization. ) Ipact of Occupancy Density: We also analyze the ipact of occupancy density on the localization accuracy of WinIPS. Fig. 1 deonstrates the functionality of each zone in the testbed. Our testbed is a real ulti-functional lab, that includes one undergraduate student office (for occupants), CDF WKNN STI-WKNN Localization error () Fig. 1. Cuulative distribution of localization error with (STI-WKNN) and without (WKNN) to copensate device heterogeneity. TABLE VI COMPARISON OF LOCALIZATION ERROR IN TERMS OF MEAN (ē LA ) AND STANDARD DEVIATION (σ LA ) BETWEEN DIFFERENT MOBILE DEVICES. Mobile device ē LA () σ LA () iphone Galaxy S Nexus ipad Air Mi one workplace for graduate students (for 7 occupants), one UAV testbed (open space), one It includes one undergraduate student office (for occupants), one workplace for graduate students (for 7 occupants), one Unanned Aerial Vehicle (UAV) testbed (open space), one workplace for undergraduate students (for 1 occupants), and one graduate student office (for occupants). As shown in the Fig. 1, the graduate student office is the ost crowded area in the lab, where the occupancy density is.7 p/. The functionality of the UAV testbed is testing the perforance of UAV so it is usually epty, which has the lowest occupancy density within the lab. Table VII elaborates the zone size, nuber of coon occupants, occupancy density and the ean localization error in each zone. As presented in Table VII, the localization accuracy in low occupancy density area is slightly better in the high occupancy density area. The ean localization error in the graduate student office is the largest (1.) and it is Fig. 1. The functionality of each zone in the testbed (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

12 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 1 TABLE VII COMPARISON OF LOCALIZATION ERROR IN EACH ZONE WITH DIFFERENT OCCUPANCY DENSITY. Zone ID Zone area ( ) No. of occupants Occupancy density (p/ ) Mean localization error () UAV testbed Workplace for graduate students Workplace for undergraduate students Undergraduate student office Graduate student office the ost crowded area in the lab, while the ean localization error in the UAV testbed is the iniu (1.7) which has the lowest occupancy density. Thus, the experiental results indicate that higher occupancy density ay affect the localization accuracy because the oveent of occupants interfere the signal propagation paths and the ultipath coponents, which contribute to higher uctuations of received signals. Potential solution to overcoe this issue is to add additional WiFi routers in the crowded area to provide ore RSSI easureents and features so that the localization accuracy can still be guaranteed. Furtherore, the localization accuracy in this can be further iproved by optiizing the placeent of APs in our previous work [3]. V. CONCLUSION In this paper, we proposed, WinIPS, a WiFi-based nonintrusive IPS that enables autoatic online radio ap construction and adaptation for calibration-free indoor localization. For RSS data acquisition, we developed WinSMS, a novel intelligent wireless syste that can capture data packets transitted in the existing WiFi traffic and extract the RSS and MAC addresses of both APs and MDs in a nonintrusive anner. We leverage APs as natural online reference points for online radio ap construction and adaptation. To construct a ore fine-grained radio ap, we further proposed the PSFM-GPR, a reliable regression technique dedicated to predicting RSS on VRPs which can well capture the nonunifor RSS distribution over coplex indoor environents. The online radio ap adapts better and is ore robust to environental dynaics than traditional offline calibrated RSS database since it keeps updated with new easureents. To alleviate the device heterogeneity issue between AP and MD, we introduced STI-WKNN, which copares the shapes of RSS vectors between RSS readings of MDs to online RSS fingerprint database rather than to raw RSS values. Extensive experients have been carried out over six onths to validate the effectiveness of WinIPS in a real-world ulti-functional office. The experiental results show that the PSFM-GPR achieves a. db average RSS estiation error and a 1.71 average localization accuracy, which outperfors existing approaches. In suary, WinIPS overcoes the bottlenecks of WiFi-based IPS, aking it proising for large-scale practical ipleentations. REFERENCES [1] D. Lyberopoulos, J. Liu, X. Yang, R. R. Choudhury, V. Handziski, and S. Sen, A realistic evaluation and coparison of indoor location technologies: experiences and lessons learned, in Proceedings of the ACM International Conference on Inforation Processing in Sensor Networks, pp. 17 1, 1. [] Z. Chen, Q. Zhu, and Y. C. Soh, Sartphone inertial sensor-based indoor localization and tracking with ibeacon corrections, IEEE Transactions on Industrial Inforatics, vol. 1, no., pp. 1 1, 1. [3] S. Lee, B. Ki, H. Ki, R. Ha, and H. Cha, Inertial sensor-based indoor pedestrian localization with iniu.1. a configuration, IEEE Transactions on Industrial Inforatics, vol. 7, no. 3, pp., 11. [] P. Bahl and V. N. Padanabhan, Radar: An in-building rf-based user location and tracking syste, in Proceedings of the IEEE Annual Joint Conference of the Coputer and Counications Societies, vol., pp. 77 7,. [] F. Zhao, H. Luo, X. Zhao, Z. Pang, and H. Park, Hyfi: Hybrid floor identification based on wireless fingerprinting and baroetric pressure, IEEE Transactions on Industrial Inforatics, vol. 13, no. 1, pp , 17. [] J. M. Pak, C. K. Ahn, Y. S. Shaliy, and M. T. Li, Iproving reliability of particle filter-based localization in wireless sensor networks via hybrid particle/fir filtering, IEEE Transactions on Industrial Inforatics, vol. 11, no., pp. 1 1, 1. [7] H. Zou, X. Lu, H. Jiang, and L. Xie, A fast and precise indoor localization algorith based on an online sequential extree learning achine, Sensors, vol. 1, no. 1, pp. 1 1, 1. [] H. Li, L.-C. Kung, J. C. Hou, and H. Luo, Zero-configuration indoor localization over ieee.11 wireless infrastructure, Wireless Networks, vol. 1, no., pp., 1. [] M. M. Atia, A. Noureldin, and M. J. Korenberg, Dynaic onlinecalibrated radio aps for indoor positioning in wireless local area networks, IEEE Transactions on Mobile Coputing, vol. 1, no., pp , 13. [1] L. M. Ni, Y. Liu, Y. C. Lau, and A. P. Patil, Landarc: indoor location sensing using active rfid, Wireless Networks, vol. 1, no., pp ,. [11] P. Krishnan, A. Krishnakuar, W.-H. Ju, C. Mallows, and S. Gat, A syste for lease: Location estiation assisted by stationary eitters for indoor rf wireless networks, in Proceedings of the IEEE Annual Joint Conference of the Coputer and Counications Societies, vol., pp ,. [1] J. Yin, Q. Yang, and L. M. Ni, Learning adaptive teporal radio aps for signal-strength-based location estiation, IEEE Transactions on Mobile Coputing, vol. 7, no. 7, pp. 3,. [13] S. Sorour, Y. Lostanlen, S. Valaee, and K. Majeed, Joint indoor localization and radio ap construction with liited deployent load, IEEE Transactions on Mobile Coputing, vol. 1, no., pp , 1. [1] A. Rai, K. K. Chintalapudi, V. N. Padanabhan, and R. Sen, Zee: zeroeffort crowdsourcing for indoor localization, in Proceedings of the ACM Annual International Conference on Mobile Coputing and Networking, pp. 3 3, 1. [1] G. Shen, Z. Chen, P. Zhang, T. Moscibroda, and Y. Zhang, Walkiearkie: Indoor pathway apping ade easy, in Proceedings of the USENIX conference on Networked Systes Design and Ipleentation, pp., 13. [1] H. Zou, B. Huang, X. Lu, H. Jiang, and L. Xie, A robust indoor positioning syste based on the procrustes analysis and weighted extree learning achine, IEEE Transactions on Wireless Counications, vol. 1, no., pp. 1 1, 1. [17] Y. Du, D. Yang, and C. Xiu, A novel ethod for constructing a wifi positioning syste with efficient anpower, Sensors, vol. 1, no., pp. 31, 1. [1] T. Roos, P. Myllyäki, H. Tirri, P. Misikangas, and J. Sievänen, A probabilistic approach to wlan user location estiation, International Journal of Wireless Inforation Networks, vol., no. 3, pp. 1 1, (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.

13 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI 1.11/TWC , IEEE Transactions on Wireless Counications 13 [1] X. Chai and Q. Yang, Reducing the calibration effort for probabilistic indoor location estiation, IEEE Transactions on Mobile Coputing, vol., no., pp., 7. [] B. Ferris, D. Haehnel, and D. Fox, Gaussian processes for signal strength-based location estiation, in Proceeding of Robotics: Science and Systes,. [1] A. M. Hossain and W.-S. Soh, Craer-rao bound analysis of localization using signal strength difference as location fingerprint, in Proceedings of the IEEE Annual Joint Conference of the Coputer and Counications Societies, pp. 1, 1. [] N. Bargshady, N. A. Alsindi, K. Pahlavan, Y. Ye, and F. O. Akgul, Bounds on perforance of hybrid wifi-uwb cooperative rf localization for robotic applications, in Proceedings of the IEEE International Syposiu on Personal, Indoor and Mobile Radio Counications Workshops, pp. 77, 1. [3] S.-S. Jan, S.-J. Yeh, and Y.-W. Liu, Received signal strength database interpolation by kriging for a wi-fi indoor positioning syste, Sensors, vol. 1, no., pp , 1. [] S. J. Pan, J. T. Kwok, Q. Yang, and J. J. Pan, Adaptive localization in a dynaic wifi environent through ulti-view learning, in Proceedings of the AAAI Conference on Artificial Intelligence, pp , 7. [] OpenWrt. [] Libpcap. [7] C. K. Willias and C. E. Rasussen, Gaussian processes for achine learning, MIT Press,. [] S. Kuar, R. M. Hegde, and N. Trigoni, Gaussian process regression for fingerprinting based localization, Ad Hoc Networks, vol. 1, pp. 1 1, 1. [] Y. Qi and H. Kobayashi, On relation aong tie delay and signal strength based geolocation ethods, in Proceedings of the IEEE Global Telecounications Conference, vol. 7, pp. 7 3, 3. [3] A. Mahtab Hossain, Y. Jin, W.-S. Soh, and H. N. Van, Ssd: A robust rf location fingerprint addressing obile devices heterogeneity, IEEE Transactions on Mobile Coputing, vol. 1, no. 1, pp. 77, 13. [31] J. G. Park, D. Curtis, S. Teller, and J. Ledlie, Iplications of device diversity for organic localization, in Proceedings of the IEEE Annual Joint Conference of the Coputer and Counications Societies, pp , 11. [3] J. C. Gower, Generalized procrustes analysis, Psychoetrika, vol., no. 1, pp. 33 1, 17. [33] K. Kaearungsi and P. Krishnaurthy, Analysis of wlans received signal strength indication for indoor location fingerprinting, Pervasive and obile coputing, vol., no., pp. 31, 1. [3] M. Jin, R. Jia, and C. Spanos, Apec: Auto planner for efficient configuration of indoor positioning syste, in Proceedings of the International Conference on Mobile Ubiquitous Coputing, Systes, Services and Technologies, pp. 1 17, 1. and Internet of Things. Han Zou received the B.Eng. (First Class Honors) and Ph.D. degrees in Electrical and Electronic Engineering fro the Nanyang Technological University, Singapore, in 1 and 1, respectively. He is currently a Postdoctoral Scholar with the Departent of Electrical Engineering and Coputer Sciences at the University of California, Berkeley, CA, USA. His research interests include ubiquitous coputing, statistical learning, signal processing and data analytics with applications in occupancy sensing, indoor localization, cyber-physical systes, sart buildings Hao Jiang obtained the B.E and Ph.D. degrees fro the School of Inforation Science and Engineering, Xiaen University, China, in and 13, respectively. He is currently an Associate Professor in the College of Electrical Engineering and Autoation, Fuzhou University, Fuzhou, China. His research interests lie in localization syste, sensor network, fiber optic sensor, and evolutionary algorith. Lihua Xie received the B.E. and M.E. degrees in electrical engineering fro Nanjing University of Science and Technology in 13 and 1, respectively, and the Ph.D. degree in electrical engineering fro the University of Newcastle, Australia, in 1. Since 1, he has been with the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, where he is currently a professor and served as the Head of Division of Control and Instruentation fro July 11 to June 1. He held teaching appointents in the Departent of Autoatic Control, Nanjing University of Science and Technology fro 1 to 1 and Changjiang Visiting Professorship with South China University of Technology fro to 11. Dr Xie s research interests include robust control and estiation, networked control systes, ulti-agent networks, and unanned systes. He has served as an editor of IET Book Series in Control and an Associate Editor of a nuber of journals including IEEE Transactions on Autoatic Control, Autoatica, IEEE Transactions on Control Systes Technology, and IEEE Transactions on Circuits and Systes-II. Dr Xie is a Fellow of IEEE and Fellow of IFAC. Costas J. Spanos received the EE Diploa fro the National Technical University of Athens, Athens, Greece, in 1, and the M.S. and Ph.D. degrees in electrical and coputer engineering fro Carnegie Mellon University, Pittsburgh, PA, USA, in 11 and 1, respectively. In 1, he joined the Faculty at the Departent of Electrical Engineering and Coputer Sciences, University of California, Berkeley, CA, USA. He has served as the Director of the Berkeley Microlab, the Associate Dean for Research in the College of Engineering, and the Chair of the Departent of Electrical Engineering and Coputer Sciences. He works on statistical analysis in the design and fabrication of integrated circuits, and on novel sensors and coputer-aided techniques in seiconductor anufacturing. He also works on statistical dataining techniques for energy efciency applications. He has participated in two successful startup copanies: Tibre Tech (acquired by Tokyo Electron) and OnWafer Technologies (acquired by KLA-Tencor). He is the Director of the Center of Inforation Technology Research in the Interest of Society and the Chief Technical Officer for the Berkeley Educational Alliance for Research in Singapore. Ming Jin is currently pursuing the Ph.D. degree in electrical engineering and coputer science with the University of California, Berkeley. His current research interests include data-efficient algoriths for iproving cyber-physical huan syste efficiency and resilience. He was the recipient of the Siebel scholarship (1), the Best Paper Award at the International Conference on Mobile and Ubiquitous Systes: Coputing, Networking and Services (1), the Best Paper Award at the International Conference on Mobile Ubiquitous Coputing, Systes, Services and Technologies (1), the Electronic and Coputer Engineering Departent Scholarship (1), the School of Engineering Scholarship (1), and the University Scholarship at the Hong Kong University of Science and Technology (1) (c) 17 IEEE. Personal use is peritted, but republication/redistribution requires IEEE perission. See for ore inforation.