A novel method of adaptive weighted K-nearest ... - SAGE Journals

Download PDF
5 downloads 0 Views 2MB Size Report
Comment
Indoor positioning, adaptive weighted, K-nearest neighbor, orientation fingerprint database, fuzzy c-means, user's orienta- tion, hybrid distance ... Monitoring, China University of Mining and Technology, 1 Daxue Road,. Xuzhou ... Cinefra22 studies the impacts of the human body on ...... Fang SH, Wang CH, Huang TY, et al.
Research Article

A novel method of adaptive weighted K-nearest neighbor fingerprint indoor positioning considering user’s orientation

International Journal of Distributed Sensor Networks 2018, Vol. 14(6) Ó The Author(s) 2018 DOI: 10.1177/1550147718785885 journals.sagepub.com/home/dsn

Jingxue Bi1,2 , Yunjia Wang1,2, Xin Li2, Hongji Cao2, Hongxia Qi2 and Yongkang Wang2

Abstract There are many factors affecting Wi-Fi signal in indoor environment, among which the human body has an important impact. And, its characteristic is related to the user’s orientation. To eliminate positioning errors caused by user’s human body and improve positioning accuracy, this study puts forward an adaptive weighted K-nearest neighbor fingerprint positioning method considering the user’s orientation. First, the orientation fingerprint database model is proposed, which includes the position, orientation, and the sequence of mean received signal strength indicator at each reference point. Second, the fuzzy c-means algorithm is used to cluster orientation fingerprint database taking the hybrid distance of the signal domain and position domain as the clustering feature. Finally, the proposed adaptive algorithm is developed to select K-reference points by matching operation, to remove the reference points with larger signal-domain distances, minimum and maximum coordinate values, and calculate the weighted mean coordinates of the remaining reference points for positioning results. The experimental results show that the average error decreases by 0.7 m, and the root mean square error decreases to about 1.3 m by the proposed technique. And, we conclude that the proposed adaptive weighted K-nearest neighbor fingerprint positioning method can improve positioning accuracy. Keywords Indoor positioning, adaptive weighted, K-nearest neighbor, orientation fingerprint database, fuzzy c-means, user’s orientation, hybrid distance

Date received: 22 November 2017; accepted: 29 May 2018 Handling Editor: Olivier Berder

Introduction The indoor positioning technology is a research focused on navigation and location-based services (LBS) and has attracted extensive attentions of research institutions, universities, and enterprises. ABI Research estimates that the commercial value of indoor locationbased services (ILBS) will be worth up to 10 billion dollars by the end of 2020.1 Nowadays, there are a large number of applications for ILBS in market. And, researchers have carried out a large number of studies on indoor positioning technologies and developed the

corresponding indoor positioning systems, such as the infrared,2 ultrasonic3 or sound,4 radio frequency 1

NASG Key Laboratory of Land Environment and Disaster Monitoring, China University of Mining and Technology, Xuzhou, China 2 School of Environmental Science and Spatial Informatics, China University of Mining and Technology, Xuzhou, China Corresponding author: Yunjia Wang, NASG Key Laboratory of Land Environment and Disaster Monitoring, China University of Mining and Technology, 1 Daxue Road, Xuzhou 221116, Jiangsu, China. Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License (http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/ open-access-at-sage).

2 identity (RFID),5,6 ZigBee,7 Wi-Fi,8,9 Bluetooth,10 Micro Electro Mechanical Systems (MEMS)11sensors, Ultra-Wideband (UWB),12 geomagnetic field,13,14 visible light,15 computer vision,16 and Pseudolites (PL).17 However, it is rather challenging to widely spread applications for most indoor positioning systems due to their poor ubiquity, on one hand, positioning systems (e.g. UWB, PL, RFID, ultrasonic, and infrared) that need special equipment are not yet integrated on the mobile terminal, on the other hand, the system using computer vision subjects to the mobile terminal’s slightly short battery life and poor computational capacity. While the wireless local area network has been widely deployed in public places, for example, the airports, shopping malls, hotels, campuses, and all smartphones integrate Wi-Fi modules. Therefore, Wi-Fi has unique advantages in ubiquitous positioning without installing additional equipment. Indoor environment is usually complex; Wi-Fi signal is influenced by presence of pedestrians, furniture, walls, and environmental changes. The phenomena of refraction, reflection, scattering, and diffraction occur during signal propagation from the transmitter to the receiver. As a result, signals arrived at the receiver are usually superposition effect of multiple signals, which are the line-of-sight (LOS) signals and non-line-of-sight (NLOS) signals under the influence of the multipath effect. Traditional indoor localization relies on the range-based techniques, which require LOS range measurement. Such positioning techniques do not perform well in indoor scenes with obstacles and room partitions. Without assuming LOS and knowing positions of access points (APs), Wi-Fi fingerprint positioning method, a process of signal matching operation, has become the simplest and most effective approach of indoor positioning. Wi-Fi fingerprint positioning method usually contains two stages: offline stage and online stage. In the offline stage, a site survey is conducted to collect the sequences of received signal strength indicator (RSSI) from all detected APs at many different reference points (RPs) with known positions. Through simple data processing and clustering analysis, the mapping relationship between the spatial position and the signal fingerprint is built at each RP. In the online stage, the collected RSSI sequences are exploited to match with the fingerprint database, to select some neighbors, the set of RPs whose similarity distances are minimum, for estimating the target position. In recent years, more attention has been paid to improve the positioning accuracy by fusion of Wi-Fi and other sensors,18–21 while the researches on Wi-Fi fingerprint positioning algorithm are relatively fewer. Cinefra22 studies the impacts of the human body on Wi-Fi RSSI and adopts the vector of mean RSSI in multiple directions to reduce the deviations caused by

International Journal of Distributed Sensor Networks the human body. Obviously, such method has not effectively explored the variations of RSSIs around the human body. The application of clustering algorithm for Wi-Fi fingerprint positioning could largely narrow search scope down, improving the computational efficiency and avoiding larger positioning errors. Altintas and Serif23 and Zhou and Fan24 adopted the K-means and fuzzy c-means (FCM) clustering algorithms to cluster RPs, respectively. Both of them took the signaldomain distance as the clustering feature, causing a not actual distribution of RPs. Li et al.25 conducted clustering analysis based on the mixing distance, which is the fusion of the signal-domain and position-domain distances, but the clustering results exploited by affinity propagation clustering (APC) algorithm are not consistent with the real indoor situations. With respect to weighted calculation process, Li et al.26 have found that when K is 5, the positioning results are better by K-nearest neighbor (KNN) or weighted K-nearest neighbor (WKNN) arithmetic. In fact, indoor scene is a complex, dynamic, and changing environment. Assuming the fixed value of K may decrease the positioning accuracy. Shin et al.27 have proposed an enhanced weighted K-nearest neighbor (EWKNN) technique to filtrate several RPs for calculating weighted coordinates, but it ignored the real spatial distribution of RPs and may lead to localization errors. Therefore, this study introduces the user’s orientation to build the orientation fingerprint database (OFPD) considering the variations of RSSIs around human body. In offline stage, the FCM algorithm is adopted to cluster for OFPD based on the hybrid distance, which is the fusion of signal-domain and position-domain distances; in the online stage, when the user requests localization, the OFPD is selected based on the user’s orientation and used for further matching operation. The proposed adaptive weighted K-nearest neighbor (AWKNN) technique could select K RPs, remove the RPs with larger signal-domain distances, minimum and maximum coordinate values, and calculate the weighted mean coordinates. This technique can make the remaining RPs more concentrated and improve positioning accuracy. This article is organized as follows. Section ‘‘The impacts of user’s orientation on Wi-Fi RSSI’’ describes the impacts of user’s orientation on Wi-Fi RSSI. In section ‘‘The proposed technique,’’ we propose the AWKNN method considering the user’s orientation. This section also involves the establishment of OFPD model and FCM clustering taking into account user’s orientation based on the hybrid distance. Section ‘‘The experiment and analysis’’ describes the experimental testbed and evaluates the performance of proposed method. We conclude the article in section ‘‘Conclusion.’’

Bi et al.

3

Figure 1. RSSI distribution under the impacts of user’s orientation: (a) RSSI collection from a certain AP in clockwise direction, orientation 1, orientation 2, orientation 3 and orientation 4 and (b) the corresponding RSSI distribution in four directions, which can reflect the impacts of user’s orientation.

The impacts of user’s orientation on Wi-Fi RSSI Some buildings are irregular with complicated interior building structures. People easily get lost in the complex indoor environment, and they are more sensitive to their front and back, right and left. Their movements are usually along the building or its vertical direction, so it is necessary to study the impacts of user’s orientation on Wi-Fi RSSI. Almost all APs are installed along buildings. And, it seems that the signal is blocked more seriously when the human body faces these four directions. So, four orientations are chose to study the influence on RSSI. As shown in Figure 1(a), a test of RSSI distribution around the human body is conducted. In the test, Wi-Fi signals are collected in the clockwise direction at the location of 1.5 m away from an AP. The sampling frequency is 1 Hz and the period is 2 min. Orient 1 faces the AP and Orient 3 is the opposite direction. Figure 1(b) shows the test results. The results of Orient 1 and Orient 2 are consistent, and their differences are small. But the signal attenuation is large in Orient 3. Due to the shelter of the human body, the maximum RSSI difference between Orient 3 and Orient 1 could reach 15 dBm. While the transmitted power level of a civil AP is around 20 dBm, the difference can be seemed as big. Compared to Orient 1, Orient 4 has smaller signal attenuation, because the right arm could block signals during the signal acquisition. Above all, the human body has significant effects on Wi-Fi RSSI, and the RSSI distribution is related to the user’s orientation. Therefore, the user’s orientation should be taken into account for Wi-Fi fingerprint positioning.

Orient 1, Orient 2, Orient 3, and Orient 4 refer to the angle scopes from 0° to 90°, 90° to 180°, 180° to 270°, and 270° to 360°, respectively.

The proposed technique This section presents the proposed Wi-Fi OFPD model, the FCM algorithm based on the hybrid distance considering user’s orientation, and AWKNN algorithm.

Wi-Fi OFPD model The traditional Wi-Fi fingerprint positioning method uses a common fingerprint database (CFPD) model, which takes the vector of mean RSSI values in multiple directions as signal features and ignores the deviations of RSSI values caused by the user’s orientation. This article puts forward a Wi-Fi OFPD model considering the user’s orientation as shown in Table 1. Where id refers to the record number, pid refers to the point number of RP, x and y refer to the coordinates of RP, AP MAC refers to the media access control address of AP, and RSSI refers to the corresponding RSSI value. In offline stage, the Wi-Fi signals are collected according to different user’s orientations using the Wi-Fi Scanner software shown in Figure 2. Subsequently, the OiFPD (i = 1, ..., 4) including the spatial position, user’s orientation, and Wi-Fi fingerprint are established. Where the O1FPD refers to the fingerprint database in Orient 1. In online stage, the user’s orientation should be determined first and then the OFPD in same angle scope could be selected for further matching operation.

4

International Journal of Distributed Sensor Networks

Table 1. The orientation fingerprint database, for example, the O1FPD. id

pid

x (m)

y (m)

Orientation (°)

AP MAC

RSSI (dBm)

1 2 .. 10 ..

1 1

0.0 0.0

0.0 0.0

23.5 23.5

6c:e8:73:91:96:d0 6c:e8:73:91:96:a1

268 247

1

0.0

0.0

23.5

6c:e8:73:90:97:d5

263

AP: access point; MAC: media access control; RSSI: received signal strength indicator.

and the cluster corresponding to the minimum signaldomain distance is selected for the next Wi-Fi fingerprint matching operation. The Euclidean distance between two sets of Wi-Fi signals is taken as the signal-domain distance. In a complex indoor environment, it is hard to ensure that every location could receive Wi-Fi signals with same number of APs. So, the number of APs is introduced to improve the signal-domain distance, as shown in equation (1), where dsignew (li , lj ) is the signal-domain distance between li and lj ; li and lj refer to the ith and jth location, respectively; RSSi, k is the RSSI of the kth AP at li and m is the amount of same detected APs at two locations dsignew



sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi m  2 1X li , lj = RSSi, k  RSSj, k m k=1 

ð1Þ

The position-domain distance dgeo (li , lj ) between two locations could be expressed as follows   qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  2  2 xi  xj + yi  yj dgeo li , lj =

Figure 2. Wi-Fi Scanner application allows researcher to collect Wi-Fi signal fingerprint at RPs and test points.

The FCM clustering based on hybrid distance considering the user’s orientation The FCM clustering based on the hybrid distance considering the user’s orientation is applied in offline stage. The feature of the adopted FCM clustering algorithm is the hybrid distance, which is the fusion of signal-domain and position-domain distances. The clustering results could be used for clustering recognition in online stage. In online stage, the OFPD is selected according to the user’s orientation, for example, the O1FPD. The signal-domain distances between the test point (TP) and all clustering centers of the O1FPD are calculated,

ð2Þ

The clustering results based on the signal-domain distance could not reflect the exact spatial relationship of RPs. Therefore, the hybrid distance, the fusion of signal-domain and position-domain distances, should be taken as the clustering feature in clustering analysis. Because of their different metrics, signal-domain and position-domain distances need to be normalized for further application. The hybrid distance dhyb (li , lj ) could be expressed as follows         dhyb li , lj = Nor dsignew li , lj 3 Nor dgeo li , lj ð3Þ

where Nor() is the normalized function, we choose the min-max normalization method, which maps the distance values to a uniform range. In practical application, we set the range to be [0.01, 1]. This process can keep the impacts of signal-domain and position-domain distances equal. Because any one of the normalized distances is set as 0, the other one is of no effect.

Bi et al.

5

Algorithm. Adaptive weighted K-nearest neighbor algorithm. Input: the orientation fingerprint database (OFPD), test point (TP) and user’s orientation (O) Output: estimated coordinates 1: if O2Orient1 then 2: OFPD=O1FPD 3. else if O2Orient2 then 4: OFPD=O2FPD 5: else if O2Orient3 then 6: OFPD=O3FPD 7: else if O2Orient4 then 8: OFPD=O4FPD 9: end if 10: Nc=Num(C), C denotes as clusters in the selected OFPD ! 2 11: for i=1 to Nc do m  P RSSjCC  RSSjTP 12: dtp cc ðiÞ = sqrt m1 j=1 13: end for 14: (d, I)=sort(dtp cc ), Csel=C(I(1)) 15: NRP=Num(Csel), denotes as the number of RPs in Csel 16: if NRP.10 then 17: K=10 18: else 19: K= NRP 20: end if 2  21: for j=1 to NRP do  P m  RSSjRP  RSSjTP 22: dsignew ð jÞ = sqrt m1 i=1 23: end for 24: (d, I)=sort(dsig-new), d1=dsig-new(I(1)), CK=Csel(I(1:K))   K K  2 P P 1 25: davg = K1 , dthr=d1 + 23dsig-s dsignew ðiÞ, dsigs = sqrt K1 dsignew ðiÞ  davg i=1

i=1

26: for k=1 to K 27: if dsig-new(k). dthr then 28: remove the kth RP in CK 29: end if 30: end for Nre P 31: Nre=Num(CK), j xjavg = N1re jxi j, i=1   2  N re P , 32: j xjs = sqrt Nre11 jxðiÞj  jxjavg i=1

33: for i=1 to Nre do 34: Sumabs ðiÞ = jxi j + jyi j   35: end for Nre Nre  2 P P ðSumabs ðiÞÞ, Sums = sqrt Nre11 Sumabs (i)  Sumavg 36: Sumavg = N1re i=1

i=1

37: for i=1 to Nre do 38: if |x(i)|\|x|avg-|x|s or |x(i)|.|x|avg + |x|sthen 39: remove the ith RP in CK 40: else if |y(i)|\|y|avg-|y|s or |y(i)|.|y|avg + |y|sthen 41: remove the ith RP in CK 42: else if |x(i)| + |y(i)|\Sumavg-Sums or |x(i)| + |y(i)|.Sumavg + Sums then 43: remove the ith RP in CK 44: end if 45: end for 46: Nls=Num(CK) denotes the number of RPs in the remaining CK, and the corresponding signal distances can be denoted as dcor  P  P Nls  Nls Nls  Nls ^ P P ^ 1 1 y 1 1 47: x = xðkÞ3 dcor ðkÞ y ðkÞ3 dcor ðkÞ dcor ðkÞ = dcor ðkÞ k=1

k=1

k=1

An AWKNN algorithm Traditional KNN and WKNN fingerprint positioning methods employ the RPs corresponding to the top K minimum signal-domain distances to estimate coordinates. These methods ignore actual spatial

k=1

distribution of RPs merely considering the impacts of signal-domain distance on positioning results. In this article, the proposed AWKNN algorithm takes both the signal-domain distance and spatial-domain distance into account. Based on certain criteria, the algorithm

6

International Journal of Distributed Sensor Networks

Figure 3. Schematic diagram of the AWKNN algorithm: (a) the three-dimensional distribution of five obtained RPs, (b) removing RP with larger signal-domain distance, (c) filtering RP with maximum coordinate absolute value, and (d) the remaining RPs with more concentrated distribution.

could select K RPs, remove the RPs with larger signaldomain distances, minimum and maximum coordinate values, and calculate the weighted mean coordinates with the remaining RPs. The specific algorithm is described as follows. Where Num() denotes that the function shows the number of available, sqrt() denotes that the function calculates the standard deviation, sort() denotes that the function sorts variables in ascending order, returning sorted variables and corresponding index, | | refers to the absolute value, subscript avg refers to the average value, subscript s refers to the standard deviation, and Sum means the sum of absolute coordinate values of the remaining RPs. The minimum signal-domain distance d1 is used to remove RPs with larger signal-domain distances, instead of the mean signal-domain distance. This is because the smaller the signal-domain distance between two points in the matching operation, the more similar they are. The RP that corresponds to the minimum signal-domain distance d1 has the highest credibility. Then, twice standard deviation is used to remove the RPs with larger signal-domain distances. The standard deviations of the absolute coordinate values are used to remove RPs with

minimum and maximum coordinate values. This step could ensure that the distribution of remaining RPs is more concentrated, avoiding larger estimation errors, and improving the positioning accuracy and stability. Assuming that five RPs P1, P2, ..., P5 are reserved, as shown in Figure 3. Figure 3(a) shows the distribution of RPs. x and y refer to the plane coordinate, respectively. The vertical direction refers to the signal-domain distance. The algorithm could be used to find and remove RPs with larger signal-domain distance, as shown in Figure 3(b), for example, the P5 in the circle. Then, P3 among the remaining RPs with the maximum coordinate value should be removed. Finally, the remaining three RPs are selected, as shown in Figure 3(d). The positioning results could be calculated by inverse distance weighted algorithm. If the coordinate value of P1 is very small, in other words, P1 deviates from P2 and P4, it will be removed by the proposed AWKNN algorithm.

The experiment and analysis This section discusses the experimental testbed and evaluates the performance of the clustering and the proposed AWKNN positioning algorithm.

Bi et al.

7

Figure 4. The layout and indoor environment of the experimental testbed.

The experimental testbed The experiment is carried out in region C, on the fourth floor of an office and teaching building, occupied by School of Environment Science and Spatial Informatics. The layout of testbed is shown in Figure 4. It includes stairs platform, corridor, and rooms. The size of corridor is 29.3 m 3 2.4 m. The length of stairs platform is 7.2 m. The size of laboratory room is 9 m 3 6.6 m. 10 TP-Link 2.4 GHz wireless APs are preinstalled at about 3 m above the corridor and stairs platform, as shown in the solid triangle. The solid dots refer to the RPs with a total of 83. The distance between a RP and its nearest neighbor RP is 1.2 m. The ‘‘fp

collect’’ button of Wi-Fi Scanner software shown in Figure 2 is used to record position, user’s orientation, and collect the corresponding Wi-Fi signal fingerprint at each RP in four different orientations. The measurement is in clockwise direction. Its sampling period is 60 s in each direction, and the sampling frequency is 1 Hz. In the fingerprint database, there are 264 positions in the corridor and stairs platform, and there are 68 positions in the room. On average, each location could receive at least eight APs’ Wi-Fi signals. To test the performance of the proposed method, 20 TPs are laid evenly in the experiment area, as shown in the hollow dots in Figure 4. The measurement at each

8 TP used by ‘‘tp collect’’ button is the same as the operation of ‘‘fp collect’’. But the sampling period is 10 s in each direction. There are 800 sets of test data in total.

Clustering results and analysis To analyze the performance of clustering of the fingerprint database in the offline stage, the K-means, FCM, and APC algorithms are used for clustering analysis based on the CFPD and OiFPD (i = 1, ..., 4), and the results are shown in Figure 5. Clustering results by the above three algorithms are successively shown from top to bottom. From left to right of Figure 5, there are five columns figures, which show clustering results by different algorithms based on CFPD, O1FPD, O2FPD, O3FPD, and O4FPD, respectively. The triangle, square, round, and star denote different clusters. Clustering results are slightly different even with the same algorithm. For instance, as shown in Figure 5(a), two clusters overlap in the area marked by rectangular solid box. And, there are different number of clusters in Figure 5(c). We can find that the clustering results may be different based on fingerprint database with different orientations. In Figure 5(c), there are four clusters of O4FPD by APC algorithm. The APC algorithm can adaptively create clusters based on the constant exchanging of reading similarities between RPs without providing initial cluster centers and counts. And, according to the building structures, decoration styles, and indoor scenes, we find that four clusters more accurately reflect the actual layout of the experimental field in Figure 4. Therefore, the number of clusters for K-means and FCM algorithm is pre-set as four. In addition, the initial clustering centers of the K-means and FCM algorithm are not randomly selected. All RPs are equally pre-divided into four clusters. The mean coordinate values and the set of mean RSSI values of all RPs in every cluster are taken as the feature of each clustering center. Figure 5(a) shows the clustering results by K-means clustering algorithm. The clustering results of CFPD are broadly consistent with those of OFPD. However, the phenomenon of cluster overlapping could be observed in the rectangular selection region in clustering results of O1FPD, which will decrease the accuracy of clustering recognition. Figure 5(b) shows the clustering results by FCM clustering algorithm. The clustering results of CFPD are also basically consistent with clustering results of OFPD. The results also exactly reflect the true indoor scenes in the experimental field. Figure 5(c) shows the clustering results by APC algorithm. The amounts of clusters are 3, 3, 3, 3, and 4 from left to right, respectively. The cluster edge changes greatly and could not truly reflect the real situations of the experiment field.

International Journal of Distributed Sensor Networks Meanwhile, the Davies–Bouldin index28,29 (1, 2)(DBI) is chosen as an effective evaluation index for clustering results. DBI is a clustering quality index, first introduced by Davies et al. in 1979. It is built to numerically express the performance of clustering tools. The smaller the DBI value, the better the clustering results. In this work, the DBI is used as a basis for assessing different clustering algorithms. Table 2 shows the DBI values of CFPD and OiFPD (i = 1, ..., 4) by K-means, FCM, and APC algorithms. In Table 2, Mean refers to the mean of DBI value and Std refers to the corresponding standard deviation. As shown in Table 2, the DBI values of CFPD and OFPD clustering are very close by K-means algorithm and FCM algorithms. The corresponding mean values and the standard derivations by K-means and FCM algorithm are also similar. The DBI values by APC algorithm are larger than those by K-means and FCM algorithm, apart from that of O3FPD. The DBI value of O4FPD by APC algorithm is more than twice that by K-means and FCM algorithms. Figure 5 and Table 2 show that the clustering by FCM algorithm has the best performance among the three clustering methods.

The evaluation of operational performance Execution time and time complexity are introduced to evaluate the performance of the proposed AWKNN algorithm compared to WKNN and EWKNN algorithms. The running platform is a computer with Intel 7 CPU, 16 GB memory and Window 10 operating system. The simulation is conducted by using of MATLAB 2016. The execution time of single localization by AWKNN algorithm is about 3.7 ms, while the execution time of single localization by WKNN and EWKNN algorithms are 2.4 and 2.8 ms, respectively. And, the time complexity of the above three algorithm is O(n), where n is the number of RPs in cluster or fingerprint database. Therefore, the proposed AWKNN algorithm has equivalent performance to WKNN and EWKNN algorithms.

The localization results and analysis The positioning experiments based on different fingerprint databases and by different algorithms are carried out, respectively. To evaluate the performance of positioning methods, the error mean (EM), root mean square error (RMSE), maximum error (ME) and 50%th cumulative distribution function error (CDFE) are used as the error statistical indexes. In the proposed AWKNN algorithm, the number of initial selected RPs is set as 10. To be fairly compared with AWKNN algorithm, different positioning tests are conducted with K

Bi et al.

Figure 5. Clustering results by K-means, FCM, and APC clustering algorithms, where triangle, square, round, and star represent different clusters: (a) the clustering results by K-means based on CFPD, O1FPD, O2FPD, O3FPD and O4FPD, respectively, (b) the clustering results by FCM based on above-mentioned fingerprint database, and (c) the clustering results by APC based on abovementioned fingerprint database.

9

10

International Journal of Distributed Sensor Networks

Table 2. DBI values by K-means, FCM, and APC algorithms. Algorithms

CFPD

O1FPD

O2FPD

O3FPD

O4FPD

Mean

Std

K-means FCM APC

0.25 0.26 0.26

0.35 0.35 0.36

0.33 0.33 0.36

0.33 0.33 0.26

0.24 0.24 0.52

0.30 0.30 0.35

0.05 0.05 0.10

DBI: Davies–Bouldin index ; FCM: fuzzy c-means; APC: affinity propagation clustering.

Figure 6. The cumulative distribution function of positioning errors by WKNN and EWKNN algorithms: (a) positioning errors based on CFPD and OFPD by WKNN algorithm and (b) positioning errors based on CFPD and OFPD by EWKNN algorithm.

values ranging from 1 to 10 by WKNN algorithm, selecting positioning results with better performance.

Results and analysis of positioning experiment based on different fingerprint databases. In this experiment, the existing WKNN and EWKNN algorithms are used for estimating results based on CFPD and OFPD, respectively. Figure 6(a) and (b) shows the cumulative distribution function of positioning errors by WKNN and EWKNN algorithms, respectively. The circles represent the positioning errors based on CFPD and the triangles represent the positioning errors based on OFPD. In Figure 6(a), the average error is minimum based on CFPD by WKNN algorithm (CFPD-WKNN) when K is equivalent to 10, while the average error is minimum based on OFPD by WKNN algorithm (OFPDWKNN) when K is 4. In Figure 6(b), the EWKNN algorithm–based CFPD (CFPD-EWKNN) and the EWKNN algorithm based on OFPD (OFPDEWKNN) have different values of K in every matching operation. Table 3 describes the statistics of positioning errors based on CFPD (CFPD-WKNN and CFPDEWKNN) and OFPD (OFPD-WKNN and OFPDEWKNN). Compared with the positioning method

based on CFPD, the values of EM, RMSE, ME, and 50%th CDFE based on OFPD decrease by 0.6, 1, 15.7, and 0.3 m. It should be noted that the value of ME is 23.3 m based on CFPD. Greater errors are introduced into localization based on CFPD. Because reflection, refraction, diffraction, and scattering may take place when Wi-Fi signals travel from transmitters to the receiver, signal enhancement or attenuation will occur in a location where the differences are small between the received Wi-Fi signal fingerprint and the signal fingerprint in the far distance in CFPD. While the OFPD includes the user’s orientation, which is useful to narrow the positioning search region and avoid larger positioning errors. As seen from Figure 6 and Table 3, the performance of the positioning method based on OFPD is superior to the CFPD-based positioning method.

Results and analysis of positioning experiment by different algorithms. First, the WKNN algorithm based on OFPD (OFPD-WKNN) is used for positioning test in different indoor scenes, and the statistics of errors are shown in Table 4. When K = 4, the minimum EM is 2.0 m in the corridor and stairs platform; when K is 5, the minimum EM is 2.8 m in the laboratory. There are

Bi et al.

11

Table 3. Statistics of positioning errors based on CFPD and OFPD (unit: m). Fingerprint type

EM

RMSE

ME

50%th CDFE

CFPD OFPD

2.9 2.3

2.4 1.4

23.3 7.6

2.4 2.1

EM: error mean; RMSE: root mean square error; ME: maximum error; CDFE: cumulative distribution function error; CFPD: common fingerprint database; OFPD: orientation fingerprint database.

Table 4. Statistics of positioning errors in different indoor scenes by OFPD-WKNN technique (unit: m). Indoor scenes

K value

EM

RMSE

ME

50%th CDFE

Corridor and stair platform Laboratory

4 5

2.0 2.8

1.3 1.4

5.7 6.5

1.8 2.7

OFPD: orientation fingerprint database; WKNN: weighted K-nearest neighbor; EM: error mean; RMSE: root mean square error; ME: maximum error; CDFE: cumulative distribution function error.

large differences in RMSE, ME, and 50%th CDFE under these two different indoor scenes. Therefore, there are errors for WKNN algorithm applying the fixed value of K. Then, the positioning experiments are conducted by the WKNN, EWKNN, and the proposed OFPDAWKNN algorithms, respectively. The cumulative probability distribution of positioning errors is shown in Figure 7. When K is 4, the minimum EM in the whole experiment could be obtained by WKNN algorithm. Table 5 gives the statistics of positioning errors by WKNN, EWKNN, and the proposed AWKNN algorithms based on OFPD (OFPD-WKNN, OFPDEWKNN, and OFPD-AWKNN). Compared with the WKNN algorithm, the minimum EM and RMSE obtained by the proposed AWKNN algorithm decrease by 1.8% and 1.9%, respectively. Compared with the EWKNN algorithm, the EM, RMSE, ME, and 50%th CDFE obtained by the proposed AWKNN algorithm decrease by 7.5%, 3.1%, 11.5%, and 6.1%, respectively. As shown in Figure 7 and Table 5, the performance of the proposed AWKNN algorithm in this article is superior to WKNN and EWKNN algorithm. Finally, the positioning experiments are conducted by CFPD-WKNN, CFPD-EWKNN, and the proposed OPFD-AWKNN technique in this article, respectively. Figure 8 shows the cumulative distribution function of positioning errors. When K is 10, the minimum EM could be obtained by CFPD-WKNN method. Table 6 shows the errors of the existing Wi-Fi fingerprint positioning methods and the proposed OFPD-AWKNN method. Compared with CFPD-WKNN, the EM, RMSE, ME, and 50%th CDFE obtained by OFPDAWKNN method decrease by 0.6, 0.9, 14.8, and 0.3 m, respectively. Compared with the CFPD-EWKNN

Figure 7. The cumulative distribution function of positioning errors by three OFPD-based algorithms.

Figure 8. The cumulative distribution function of positioning errors of the existing algorithms and the proposed algorithm.

12

International Journal of Distributed Sensor Networks

Table 5. Statistics of positioning errors by OFPD-based three algorithms (unit: m). Methods

EM

RMSE

ME

50%th CDFE

OFPD-WKNN (K = 4) OFPD-EWKNN OFPD-AWKNN

2.2 2.4 2.2

1.4 1.4 1.4

6.0 7.6 6.7

2.0 2.2 2.0

OFPD: orientation fingerprint database; EM: error mean; RMSE: root mean square error; ME: maximum error; CDFE: cumulative distribution function error; WKNN: weighted K-nearest neighbor; EWKNN: enhanced weighted K-nearest neighbor; AWKNN: adaptive weighted K-nearest neighbor.

Table 6. Statistics of positioning errors of the existing algorithms and the proposed algorithm (unit: m). Methods

EM

RMSE

ME

50%th CDFE

CFPD-WKNN (K = 10) CFPD-EWKNN OFPD-AWKNN

2.8 3.1 2.2

2.3 2.6 1.4

21.5 23.3 6.7

2.3 2.4 2.0

OFPD: orientation fingerprint database; EM: error mean; RMSE: root mean square error; ME: maximum error; CDFE: cumulative distribution function error; WKNN: weighted K-nearest neighbor; EWKNN: enhanced weighted K-nearest neighbor; AWKNN: adaptive weighted K-nearest neighbor.

method, the values of EM, RMSE, ME, and 50%th CDFE by OFPD-AWKNN method decrease by 0.9, 1.2, 16.6, and 0.4 m, respectively. Compared with the existing methods, the values of EM, RMSE, ME, and 50%th CDFE by OFPD-AWKNN method decrease by 0.7, 1.0, 15.7, and 0.3 m on average, respectively. In this article, the proposed OFPD could narrow the search scope of the matching operation and avoid larger errors. The AWKNN method considers the signal-domain distance between TP and RPs and the real position distribution of RPs simultaneously. It could remove the RPs with the larger signal-domain distance, the minimum and maximum coordinate values, and improve the fingerprint positioning accuracy. The experimental results indicate that the performance of the newly proposed OFPD-AWKNN method is superior to the CFPD-WKNN and CFPD-EWKNN methods.

Conclusion This article studies the impacts of the human body on Wi-Fi RSSI. To eliminate or weaken positioning errors caused by user’s human body, user’s orientation is introduced to establish the OFPD. And, the FCM clustering based on the hybrid distance is exploited in offline stage. Moreover, an AWKNN method considering the signal-domain distance and exact position distribution of RPs is proposed. This method is suitable for the indoor scenes of corridor, stairs platform, and rooms and could improve the accuracy of Wi-Fi fingerprint positioning. The EM, RMSE, ME, and 50%th CDFE

obtained by the OFPD-AWKNN are 2.2, 1.4, 6.7, and 2.0 m, respectively. Compared with CFPD-WKNN and CFPD-EWKNN methods, the average positioning accuracy of OFPD-AWKNN method increases by 0.7 m and the RMSE decreases by 1 m on average. However, the densely distributed RPs in the whole experiment field bring great difficulties for fingerprint data acquisition. In addition, the Wi-Fi signals vary greatly in indoor environment. Therefore, the Gaussian process, interpolation, and adaptive update fingerprint technologies should be introduced for the further researches to reduce the workload of fingerprint acquisition. Declaration of conflicting interests The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Key Research and Development Program of China (grant number 2016YFB0502102), the Fundamental Research Funds for the Central Universities (grant number KYLX16_0544), the National Natural Science Foundation of China (grant number 41674030), and the China Postdoctoral Science Foundation (grant number 2016M601909).

ORCID iD Jingxue Bi

https://orcid.org/0000-0002-2964-7698

Bi et al. References 1. Connolly P and Boone D. Indoor location in retail: where is the money? Oyster Bay, NY: ABI Research, 2016. 2. Zhang Y, Chen HS and Luo Y. A novel infrared landmark indoor positioning method based on improved IMM-UKF. Appl Mech Mater 2014; 511–512: 880–885. 3. Lin ZJ, Xiong J, Guo H, et al. Beacon identification method of ultrasound indoor positioning based on FDMA. Meas Control Tech 2016; 6: 474039. 4. Huang W, Xiong Y, Li XY, et al. Swadloon: direction finding and indoor localization using acoustic signal by shaking smartphones. IEEE T Mobile Computing 2015; 14(10): 2145–2157. 5. Xi L, Jiang MY and Liao CZ. Indoor positioning system based on RFID. Appl Mech Mater 2014; 543–547: 2627–2631. 6. Ni LM, Liu Y, Lau YC, et al. LANDMARC: indoor location sensing using active RFID. IEEE Int Conf Pervasive Comput Commun 2003; 14: 2145–2157. 7. Fang SH, Wang CH, Huang TY, et al. An enhanced ZigBee indoor positioning system with an ensemble approach. IEEE Commun Lett 2012; 16(4): 564–567. 8. Walters C. Characterization of smart phone received signal strength indication for WLAN indoor positioning accuracy improvement. J Network 2014; 9(3): 1061–1065. 9. Bahl P and Padmanabhan VN. RADAR: an in-building RF-based user location and tracking system. In: Proceedings of the INFOCOM 2000 nineteenth joint conference of the IEEE computer and communications societies proceedings, IEEE; 2000. 10. Li H. Low-cost 3D Bluetooth indoor positioning with least square. Wireless Pers Commun 2014; 78(2): 1331–1344. 11. Zheng L, Zhou W, Tang W, et al. A 3D indoor positioning system based on low-cost MEMS sensors. Simul Model Pract Th 2016; 65: 45–56. 12. Garcı´ a E, Poudereux P, Herna´ndez A´, et al. A robust UWB indoor positioning system for highly complex environments. In: Proceedings of the IEEE international conference on industrial technology, Seville, 17–19 March 2015. New York: IEEE. 13. IndoorAtlas L. Ambient magnetic field-based indoor location technology: bringing the compass to the next level. Oulu: IndoorAtlas Ltd., 2012. 14. Xie H, Gu T, Tao X, et al. MaLoc: a practical magnetic fingerprinting approach to indoor localization using smartphones. In: Proceedings of the ACM international joint conference on pervasive and ubiquitous computing, Seattle, Washington, DC, 2014. 15. Hassan NU, Naeem A, Pasha MA, et al. Indoor positioning using visible LED lights: a survey. ACM Comput Surv 2015; 48(2): 20. 16. Sevrin L, Noury N, Abouchi N, et al. Characterization of a multi-user indoor positioning system based on low cost depth vision (Kinect) for monitoring human activity in a smart home. In: Proceedings of the 37th annual international conference of the IEEE engineering in medicine and biology society (EMBC), Milan, 25–29 August 2015, pp.5003–5007.

13 17. Fujii K, Yonezawa R, Sakamoto Y, et al. A combined approach of Doppler and carrier-based hyperbolic positioning with a multi-channel GPS-pseudolite for indoor localization of robots. In: Proceedings of the international conference on indoor positioning and indoor navigation, Alcala de Henares, 4–7 October 2016. New York: IEEE. 18. Rantakokko J, Rydell J, Stro¨mba¨ck P, et al. Accurate and reliable soldier and first responder indoor positioning: multisensor systems and cooperative localization. IEEE Wireless Commun 2011; 18(2): 10–18. 19. Baniukevic A, Jensen CS, Lu H, et al. Hybrid indoor positioning with Wi-Fi and Bluetooth: architecture and performance. In: Proceedings of the IEEE international conference on mobile data management, Milan, 3–6 June 2013. New York: IEEE. 20. Liu J, Chen R, Pei L, et al. A hybrid smartphone indoor positioning solution for mobile LBS. Sensors 2012; 12(12): 17208–17233. 21. Wang H, Sen S, Elgohary A, et al. No need to war-drive: unsupervised indoor localization. In: Proceedings of the international conference on mobile systems, applications, and services, Low Wood Bay, 25–29 June 2012. New York: IEEE. 22. Cinefra N. An adaptive indoor positioning system based on Bluetooth low energy RSSI, 2014, https://www.polite si.polimi.it/handle/10589/92284 23. Altintas B and Serif T. Improving RSS-based indoor positioning algorithm via K-means clustering. In: Proceedings of the wireless conference 2011—sustainable wireless technologies, Vienna, 27–29 April 2011. New York: IEEE. 24. Zhou H and Van NN. Indoor fingerprint localization based on fuzzy C-means clustering. In: Proceedings of the international conference on measuring technology & mechatronics automation, Zhangjiajie, China, 10–11 January 2014. New York: IEEE. 25. Li X, Wang J, Liu C, et al. Integrated Wi-Fi/PDR/smartphone using an adaptive system noise extended Kalman filter algorithm for indoor localization. ISPRS Int J GeoInform 2016; 5(2): 8. 26. Li B, Salter J, Dempster AG, et al. Indoor positioning techniques based on wireless LAN. In: Proceedings of the 1st IEEE international conference on wireless broadband and ultra wideband communications, Sydney, 13–16 March 2007, pp.13–16. New York: IEEE. 27. Shin B, Lee JH, Lee T, et al. Enhanced weighted K-nearest neighbor algorithm for indoor Wi-Fi positioning systems. In: Proceedings of the international conference on computing technology and information management, Seoul, 24–26 April 2012. New York: IEEE. 28. Davies DL and Bouldin DW. A cluster separation measure. IEEE T Pattern Anal 1979; PAMI-1(2): 224–227. 29. Coelho GP, Barbante CC and Boccato L, et al. Automatic feature selection for BCI: an analysis using the Davies-Bouldin index and extreme learning machines. In: Proceedings of the international joint conference on neural networks, Brisbane, QLD, Australia, 10–15 June 2012. New York: IEEE.