홈으로ArticlesAll Issue
ArticlesCOMPASS: An Active RFID-Based Real-Time Indoor Positioning System
  • Yung-Fu Hsu, Chu-Sung Cheng, and Woei-Chyn Chu*

Human-centric Computing and Information Sciences volume 12, Article number: 07 (2022)
Cite this article 1 Accesses
https://doi.org/10.22967/HCIS.2022.12.007

Abstract

Location-awareness has attracted attention over the past decades because of growing commercial interest in indoor location-based service. The k-nearest neighbor (kNN)-based positioning methods perform poorly in complex indoor environments due to ambiguity in the received signal strength indicator (RSSI). We propose a pragmatic real-time positioning framework, the Community-Optimized Measuring of Positions Associated with Sensing Signals (COMPASS), to improve positioning accuracy. This paper presents a framework to increase the probability of selecting the optimal neighbors. Our framework is implemented through an active RFID network. First, a Kalman filter (KF) is used to filter fluctuations in raw RSSIs. Second, instead of using individual tags to calculate positions, COMPASS divides the sensing area into a number of regions called “communities.” Third, a community is elected from the community chain of signal strength disparity that has the highest probability of enclosing the tracking point. Experimental results show that, compared to traditional kNN, KF-COMPASS had a 67.2% increased probability to select correct neighbors. In addition, the mean absolute error and root mean square error are 105 cm and 120 cm, which are comparable to recent studies.


Keywords

Location-Awareness Indoor Positioning, Radio-Frequency Identification (RFID), RSSI Path-Loss Model, Kalman Filter


Introduction

The Internet of Things (IoT) generated many interests in smart living over the past few decades. It refers to uniquely identifying objects (things) and delivering their information to an Internet-like structure. One of its many interesting applications is the location-based service (LBS), such as the global navigation satellite system (GNSS), local positioning system (LPS), and indoor positioning system (IPS). Global positioning system (GPS) is perhaps the most renowned GNSS widely used for reliable positioning, navigation, and timing. A key mechanism in GPS for localization is the use of time-of-arrival (ToA) information and triangulation to calculate an object’s position. Although GPS offers near-worldwide coverage, its performance degrades sharply indoors and in high-rise urban areas. Accordingly, technologies such as radio frequency identification (RFID), Bluetooth, and Wi-Fi are more favorable surrogates for accurate localization of objects within buildings [13]. Among them, RFID is also considered as a prerequisite for IoT, such as identification, object-oriented tags, cooperation with other sensors and fundamental network communication [461]. In reality, RFIDs have been developed for a wide variety of applications including supply chain management, healthcare equipment management, theft and counterfeit prevention, asset and inventory tracking [7], etc. In particular, it has wide applications in indoor tracking of care-needed people, such as children, elderlies, and patients in the blooming long-term healthcare market [8, 9].
As far as the localization strategies are concerned, the perceived signal of a radio transmitting system can be categorized into the following techniques [10]: ToA, angle-of-arrival (AoA), phase difference-of-arrival (PDoA), and received signal strength indicator (RSSI). ToA is known for having a small location error, but it requires extremely accurate time synchronization and high-cost hardware. Instead, time difference-of-arrival (TDoA) can largely improve the source-to-receiver synchronization problem by a set of hyperboloids between the source and two receivers. Alternatively, AoA requires fewer measurement units to implement; but it is vulnerable to the multipath or reflection effects. Thirdly, the PDoA not only delivers high accuracy but also detects small movement of an object, however, the multipath effect would cause tricky phase distortion. The RSSI-based technique often is limited by the fact that the realistic properties of RSSI are unsteady and vulnerable. Even so, it is highly promising and is the most commonly applied practice for location tracking because of its advantages in ubiquity, low-complexity, and easy measurement.
In RSSI-based framework, several empirical and theoretical models have been proposed for localization purposes. SpotON [11], one of the pioneering location sensing systems, principally utilized a trilateration technique according to a simple RSSI-distance relation. Later new models/algorithms were proposed to enhance its performance. Yang and Liu [12] determines an optional trilateration with the highest quality from accessible beacon nodes. Thaljaoui et al. [13] made use of the estimated distance to draw rings instead of circles around beacons to diminish the conventional problem of a noisy intersection or no intersection while trilateration. Khanh et al. [14] implemented the trilateration on Wi-Fi architecture and composed a cloudlet-based navigation. These model-based methods have convenience, fewer landmarks, and low complexity properties, but they require an accurate signal propagation model.
To improve the accuracy of position estimation, some researchers proposed the use of reference points (RPs) with known and fixed locations to characterize the sensing environment. The pattern is known as scene analysis, which refers to collecting environmental concerning features as a fingerprint by comparing online clusters with the nearest a priori feature of the dataset. The k-nearest neighbor (kNN) algorithm is one typical representative therein [1], but traditional kNN often misjudge the nearest neighbors due to the ambiguity of the RSSIs. Therefore, Han and Cho [15] reinforced kNN by choosing the best estimation error as the output from adaptive kNNs. Li et al. [16] strengthened the qualification on the selection of k candidates through building an RSS-level scaling weight on the assessment of Euclidean distance. Xu et al. [17] established kNN with a combination of Bayesian probability based on passive ultra-high frequency (UHF) RFID. The method applied a Gaussian filter to reduce the influence of RSSI fluctuation and introduced Bayesian estimation into the derivation of a target’s location. Lately, Afuosi and Zoghi [18] proposed a Wi-Fi positioning algorithm by the fingerprint clustering and a weighted differential coordinate information (WDCI-kNN). The scheme used the average of positioning error of online layer to calibrate weighted-kNN method. In General, kNN-based studies were dedicated to reforming the selecting mechanism of k neighbors and decreasing the effect of RSSI variation, it remains a challenge to choose the “correct” reference points to minimize interpolation errors.
Machine learning is another emerging solver in the positioning field that intends to learn fingerprint features and then classify [19], such as statistical learning [2022] and artificial neural network (ANN) [2325]. Lee et al. [22] combined a smart watch with Wi-Fi fingerprint based on random forest learning model. Hoang et al. [25] developed an indoor localization system through long short-term memory of recurrent neural network. Riady and Kusuma [26] used the hybrid method of ANN-based fingerprint and pedestrian dead reckoning to provide indoor positioning by the Bluetooth low energy (BLE) technology. On the whole, machine learning provides a fairly high level of accuracy due to a posteriori data mining of targeted environments. However, it has a high calibration cost that requires large dataset samples to achieve the intended generalization. Moreover, the static database is usually vulnerable to environmental dynamics rendering it costly to reconstruct the offline survey and largely increases maintenance costs and efforts.
In this paper, we propose an alternative approach in an attempt to improve indoor localization accuracy using active RFID, which we name COMPASS (Community-Optimized Measuring of Positions Associated with Sensing Signals). To prove our work, we established an active RFID network to perceive an experimental environment. The identified information can be collected by readers and transmitted to the processing server, where RSSI is first denoised by a Kalman filter (KF). The filtering process alleviates RSSI fluctuation. Subsequently, a community chain of reference tags will be introduced to determine the optimal community that has the highest likelihood to encompass the tracking tag. The process reasonably utilizes fingerprinting RSSI of spatial correlation to reduce the influence of RSSI ambiguity, which raises the probability to select the correct region. Moreover, community dimension is a time-varying variable that scales in response to the noise level of the measured RSSIs. These processes ensure that COMPASS will produce better positioning accuracy for indoor location-based services. In practice, the contributions of our proposed framework include the following:

COMPASS is ready for deployment and compatible with fingerprint-based wireless positioning systems because the concept is easy to integrate and fuse.

We propose a fingerprint of spatial correlation, i.e., the community chain. It is able to determine the optimal target’s neighbors.

The proposed COMPASS framework is an arbitrary community, say rank 2 or rank 3, independent of different noise levels.

COMPASS has higher accuracy than traditional kNN. The estimated locations have better steadiness.

Our proposals were tested in a real environment as shown in Section 4.

The remainder of this paper is organized as follows. In Section 2, we give a review of related work. In Section 3, we describe the proposed methodology including network system architecture, data processing, spatial communities and location estimation. In Section 4, we present the simulation and experimental results and validate our theoretical model for locating objects. In Section 5, we discuss the comparisons between the proposed method and other approaches. Finally, in Section 6, we draw conclusions and summarize the contribution of the present work.


Related Work

In this section, we describe literature related to our work. In light of RSSI characteristics, we explain the background, influencing factors and filtering methods towards RSSI. Subsequently, we reference several previous kNN-based positioning methodologies.

RSSI Characteristics
Considering the shadow fading inside a building or densely populated areas, an empirical model has been recognized and extensively used to address the wireless propagation loss: the so-called log-distance model [27]. In the model, the received signal strength is defined as

$PL(d)[dB]=PL(d_0)+10ηlog_{10}(\frac{d}{d_0})+Xσ,$(1)

where the power $PL(∙)$ is given in logarithmic scale over the significant distances $d, PL(d_0)$ is the power measurement at a reference distance $d_0$ by defining 1 m away from a receiver. $X_σ ~ N(0,σ^2)$ is hypothesized as zero-mean Gaussian distributed random variable with the variance $σ^2$ dB which is created by the effect of shadow fading. Here the path loss exponent (PLE) $η$ depends on a specific propagation environment that can be fitted over the received dataset at a wide range of measurement locations.
In fact, the RSSI corresponding to the sensing distance is often not a monotonous logarithmic curve in spite of showing a declining tendency. The appearance is usually caused by various factors, such as multipath, obstacles, tag orientation or moving objects among the receiver and the transmitter. In order to lessen the fluctuation of RSSI, some researchers proposed the denoising process using KF or enhanced KF. Mackey et al. [28] compared several smoothing filters and concluded that KFs offered significant localization improvement over raw RSSIs and a simple moving filter. Huang et al. [29] removed the RSSI drift using KF and supplied these steady signals into Heron-Bilateration location estimation. However, RSSI variations do not usually appear as a Gaussian distribution, which may impact the efficacy of applied KF. In view of this condition, Li et al. [23] proposed an RSSI correction method to obtain accurate RSSIs. The method aims to eliminate the unsteadiness caused by the transmitter’s emission power, and can be calibrated to an accurate RSSI distance model, denoted as Equation (2).

$PL(d)=PL(d)-ΔPL(d_0)\\ = -10ηlog_{10}(\frac{d}{d_0})+\overline{PL}(d_0) $(2)

where $\tilde{PL}(d)$ is the adjusted RSSI at a moving distance $d, PL(d)$ is the online coming RSSI, and $∆PL (d_0)=PL(d_0)-\overline{PL}(d_0)$ which $PL(d_0)$ and $\overline{PL}(d_0)$ are the online RSSI and offline averaged RSSI at a defined reference distance $d_0$.

kNN-based Approaches
One popular kNN-based approach is the LANDMARC [30], which used a set of reference tags as landmarks to calculate the object’s location. The landmarks essentially can be integrated as a featured fingerprint that carries unique IDs, known locations, and the RSSI dataset. The RSSI dataset between the reference tags and the unknown positioning targets can be matched by the Euclidean distance,

$E_l=\sqrt{\displaystyle\sum_{n=1}^N(t_n-η,n)^2,l=1,2,...,L}$(3)

where $t_n$ is the nth reader’s RSSI acquired at the target points, and $r_{l,n}$ is the RSSI of the same reader at an lth reference point in the fingerprint. N and L are the total number of readers and reference points deployed in the environment.
According to Equation (3), a number of k values are selected and recognized as the k nearest reference tags to the target. The target’s location $(x,y)$ can therefore be deduced by the weighted Euclidean distance (Euclidean-kNN),

$(x,y)=\displaystyle\sum_{i=1}^k ω_i(x_i,y_i)$(4)

where ($x_i,y_i$) is the known coordinate of the $i$th neighboring reference tag. The factor $w_i$ denotes that a smaller signal strength difference of the $i$th candidate in Equation (4) will have a larger weight in location calculation.
The shortage with this approach is that undesired candidates may be subsumed due to the ambiguity of the RSSIs. Besides, the kNN-based approach necessarily considers the layout of tag-to-tag density or the deployed distance among tags [31]. Subsequently, the NICoT approach [32] used virtual reference points of the nonlinear interpolation and a credibility mechanism to elevate estimation accuracy. NICoT reached better location accuracy than previous methods using virtual tags, but it led to larger computational load because more virtual reference tags were involved in the calculation.
On the other hand, feature matching is a considerable aspect of the improvement. Luo et al. [33] designed human-centric collaborative feedback with the combination of Wi-Fi fingerprints. The mechanism quickly self-corrected positioning likelihood by subsequent users’ positive/negative responses. Wang et al. [34] proposed a signal weighted distance (SWED)-kNN approach according to the nonlinear relationship between the RSSI and the physical distance in the propagation model. Compared to Euclidean-kNN and Manhattan-kNN, SWED-kNN attained better accuracy in light of sight environments. Xie et al. [35] proposed the improved Spearman distance to substitute for the traditional Euclidean distance. Compared to the conventional kNN method, the results were greatly improved not only in the aspect of the number of closest neighbors but also in the impact of shadowing effect. However, the approach has a higher time complexity than Euclidean distance and Pearson correlation coefficient [36]. Recently, Hunag et al. [37] adopted universal kriging algorithm to carry out fingerprint-based indoor positioning on 5G networks. The algorithm mainly estimated the weight coefficient matrix of fingerprint in the offline phase, and mapped to online kNN positioning afterwards. Wang and Park [38] presented a Wi-Fi-based positioning method that combined physical layer channel state information (CSI) and RSSI. The crucial step was to determine nearest RSSI, CSI amplitude and CSI phase between reference points and a tracking point (TP).


Methodology

In this section, we first introduce our proposed RFID network positioning architecture. When the required information was collected to the backend server, the sensing environment can be transformed into a fingerprint structure that carries the information of deployed landmarks. And afterwards the RSSI dataset in the fingerprint will be conducted through a series of processing procedures, i.e., pre-processing, filtering, community checkpoint and weighted computation to estimate the TP location. The detailed description of each measure follows below.

System Architecture
With the proof-of-concept of establishing a positioning preparation ahead of IoT infrastructure, we conceive a pragmatic wireless communication framework based on active RFID. This platform integrates a physical RFID network for a variety of uses, mobilizes information from RFID devices, and addresses challenging problems on indoor object positioning. The system architecture is depicted in Fig. 1. It consists of the RFID devices, tagged objects, a wireless network, a location database, a processing server, and available enterprise applications.

RFID devices : The operating frequency is 433 MHz, which is governed by the ISO 18000-7 and within the UHF spectrum of the industrial, scientific and medical (ISM) radio bands. In our experimental design, all readers and reference tags are fixed with known coordinates.

Tagged objects: RFID tags are affixed to the objects to be tracked. They are considered to be a single component in the RFID system.

Wireless network: To enable the wireless interconnection between readers in the RFID subsystem and the server, a wireless distribution system (WDS) is deployed. WDS allows the wireless network to be expanded using multiple access points connected to the readers under the IEEE 802.11 protocol without the need for a wired backbone.

Location database: To simplify the development and automate deployment of our RFID system, a database is used to maintain the readers’/reference tags’ positions and fingerprint information for real-time computing of the locations of the tracking objects.

Processing server: The server is responsible for computing the tagged object’s position based on the data collected from the readers. With a robust positioning algorithm described in the rest of this section, the COMPASS is able to fulfill subsequent applications in a variety of fields.

Enterprise applications: Major applications include supply chain management, asset management, animal tracking, logistics tracking, care-needed people tracking, and contact tracing of coronavirus disease 2019 (COVID-19) [39], etc.



Fig. 1. A set of 433 MHz radio frequency identification tags and readers are used to setup our wireless communication framework. With a location database and processing algorithms, the system is able to perform real-time object tracking for various enterprise applications.


Localization Algorithm
In the deployment of our RFID indoor localization system, there are $L=(I×J)$ RPs uniformly distributed in the sensing area as illustrated in Fig. 2. The objective is to calculate the positions of the TPs by the aid of RPs placed at known and fixed locations. In the time domain, we define the time series $t_τ$ that $τ$ stands for the current time point. Therefore, the signal strength vector $\overset{\small\rightharpoonup}{r}_l$ at the $t_τ$ for the lth RP of a position ($i,j$) give $\overset{\small\rightharpoonup} {r}_l(t_τ) = {φ_1(t_τ),φ_2(t_τ),…,φ_N(t_τ)}$ , where $φ_n$ is the raw RSSI perceived by the nth reader, $n ∈ [1,N]$ with N the number of readers. Meanwhile, the RSSI vector for the oth TP is defined as $\overset{\small\rightharpoonup} {r}_o(t_τ) = {ψ_1(t_τ), ψ_2(t_τ),…,ψ_N(t_τ)}$ , where $O$ is the number of tracking targets. Moreover, as far as the reference distance is concerned, we arranged $N$ tags and each tag placed one meter apart from the $N$ readers. Accordingly, the perceived RSSI vector of the reference distance at $t_τ$ can be recorded as $\overset{\small\rightharpoonup} {δ}(t_τ)={δ_1(t_τ),δ_2(t_τ),… ,δ_N(t_τ)}$ .
After building the reference topography, the distributed features of the entire sensing area in a fingerprint form becomes

(5)

where F retains the structure $f(∙)$ of each RP contained the time-varying recordings of $(ID,\overset{\small\rightharpoonup}{d} _l,\overset{\small\rightharpoonup} {r}_l, \overset{\small\rightharpoonup}{\hat{r}}_l). ID$ is the unique number with regard to the lth RP, $\overset{\small\rightharpoonup} {d}_l$ is the constant distance vector as $\overset{\small\rightharpoonup} {d}_l = {d_1,d_2,…,d_N}$ based on the physical distances of the lth RP to the numbered readers $[1,N]$, and $\overset{\small\rightharpoonup}{\hat r}_l(∙)$ is the vector of the modelized RSSIs.

Fig. 2. Illustration of the concept of a community in the proposed COMPASS algorithm. There are a total of $I×J$ reference tags uniformly deployed in a sensing area. Each community is composed of $m×m$ reference tags, 2×2 in this example, and the one that has the highest probability to encompass the tracking tag is used for the subsequent position calculation.


Pre-processing of RSSI
In order to adapt to environment dynamics, the proposed system conducts adaptive PL modeling on the fingerprint F in real-time rather than the offline mode. By this way, the adaptive model at $t_τ$ is dependent on a dataset of prior experiences which is constrained within

(6)

where $Τ$ is the time interval that allows access to the corresponding range of records, also, its maximum interval queue is equal to the sliding size $S$. Hence, the averaged RSSI vector at a reference distance within $Τ$ becomes $\overset{\small\rightharpoonup}{\overline δ}(t_τ)={\overline δ_1,\overline δ_2,…,\overline δ_N},$, and each measured RSSI in $\overset{\small\rightharpoonup} r_l(Τ)$ and $\overset{\small\rightharpoonup} r_o(Τ)$ is calibrated by

$\overset{\small\rightharpoonup}{\tilde r}_l(T)=\overset{\small\rightharpoonup}{r_l}(T)-Δδ(T)$
$\overset{\small\rightharpoonup}{\tilde r}_o(T)=\overset{\small\rightharpoonup}{r_o}(T)-Δδ(T)$ (7)

where $\overset{\small\rightharpoonup}{\tilde r}_l(T)$ and $\overset{\small\rightharpoonup}{\tilde r}_o(T)$ are the vectors of the calibrated RSSI for the $l$ th RP and the oth TP, and $∆ \overset{\small\rightharpoonup}{δ}$ can be calculated by $∆ \overset{\small\rightharpoonup}{δ}(T)$ = $\overset{\small\rightharpoonup}{δ}(T)-\overset{\small\rightharpoonup}{δ}(t_τ)$. Accordingly, the averaged RSSI of the calibrated vector for the RPs and TPs are denoted as $\overset{\small\rightharpoonup}{\overline r}_l(t_τ)$ and $\overset{\small\rightharpoonup}{\overline r}_o (t_τ)$, respectively.
In Equation (1),$PL(d)$ is proportional to a logarithmic function of a distance $d, X_σ$ has a mean 0, and the PLE $η$ is assumed to be an identical environment parameter due to large-scale fading. For these reasons, let

$B=uA+v$(8)

be the linear equation system to describe the empirical PL model on $F$, where $Β$ denotes a dataset of $\overset{\small\rightharpoonup}{\overline r}_l(t_τ)$, Α signifies a set of logarithmic function of $\overset{\small\rightharpoonup}{d}_l, u$ is the scaling value $η$, and $v$ is the overall intercept $δ$. To help computation, vectors $B$ and $A$ are cascaded to a length of $N×L$, so the series of Equation (8) becomes

(9)

In order to minimize the residual error $h_{nl}$ between the predicted values and measurements, here we apply the least square method (LSM) to solve the unknown eigenvector of $b=[u v]$ by

$H(u,v) = \sqrt{\frac{1}{(N.L)}\displaystyle\sum_{nl=1}{N.L}(β_nl-(uα_{nl}+v))^2},$(10)

where $β_{nl}=\overlineφ_{nl}, α_{nl} = d_{nl}, nl ∈ [1,NL]$, and $H$ is the root mean square error (RMSE) function as well as the signification of environmental shadowing effect. To optimize u and v, all the partial derivatives of the $H(u,v)$ function must equal to zero that yields $N×L$ equations as $∂H⁄∂u = 0$ and $∂H⁄∂v = 0$. Thus, Equation (9) can be put in a matrix form by

pyo(11)

which the desired eigenvector is calculated by $b=(Α^TA)^-1 A^TΒ$. Accordingly, the adaptive PL model at $t_τ$ can be modelized to

pyo(12)

where $\overset{\small\rightharpoonup}{\hat r}_l(t_τ)$ is the modelized vector of the lth RP and $\hat φ_n$ denotes the modelized RSSI. In the aspect of the TP, the estimated distance vector $d_o$ can be inferred as

pyo(13)

therefore, the modelized RSSI vector of the TPs $\overset{\small\rightharpoonup}{\hat r}_o$ at $t_τ$ can be obtained by

pyo(14)

The Intervention of Kalman Filter
In reality, the measured RSSIs are usually influenced by power variation and some environmental interference. KF is one effective algorithm that makes estimates of some unobserved variables based on noisy measurements. Therefore, consider a discrete-time system of the state equation with unobserved $x_τ$ at time instant $τ$ , it can be represented as

$x_r = x_{r-1}+w_{τ-1},$(15)

where $x_τ$ is the modelized RSSI value $\hat φ_n(t_τ)$ of the lth RP or $\hat ψ_n(t_τ)$ of the oth TP, and the process noise covariance matrix is given as $Q_τ = E[w_τ w_τ^T] . The observation equation can be expressed as

$Z_τ = X_τ + V_τ$(16)

where $z_τ$ is the measured RSSI value $φ_n(t_τ)$ of the lth RP or $ψ_n(t_τ)$ of the oth TP, and the observation noise covariance matrix is $R_τ=E[v_τ v_τ^T]$.
The estimation status is comprised of two significant phases, the predicted procedure, and the update procedure. The predict (a priori) procedure is modeled as,

$\hat x_{τ|τ-1} = \hat x_{τ-1|τ-1}$
$P_{τ|τ-1} = P_{τ-1|τ-1} + Q_τ$ (17)

and the update (a posteriori) procedure is modeled as,

$K_τ = P_{τ|τ-1}(p_{τ|τ-1} + R_τ)^{-1}$
$\hat x_{τ|τ-1} = \hat x_{τ|τ-1} + K_τ(Z_τ - \hat x_{τ|τ-1})$
$P_{τ|τ} = P_{τ|τ-1}(1-L_τ),$ (18)

where $\hat x_{τ|τ-1}$ is the predicted state estimation, $p_{τ|τ-1}$ is the predicted error covariance, $K_τ$ stands for the optimal Kalman gain, $\hat x_{τ|τ}$ represents the optimal state estimation, and $P_{τ|τ}$ means the updated estimation covariance.
Consequently, the RSSI vectors of Kalman filtering for RPs and TPs can be reformed to $\overset{\small\rightharpoonup}{R}_l(t_τ) = {Φ_1(t_τ),Φ_2(t_τ),…,Φ_N(t_τ)}$ and $\overset{\small\rightharpoonup}{R}_o(t_τ) = {Ψ_1(t_τ),Ψ_2(t_τ),…,Ψ_N(t_τ)}$ , respectively. The intention is to reduce the fluctuation of RSSI in F by underlying the time-varying adjustment

Community Chain
Once the RSSIs are sorted out, the comparison between $\overset{\small\rightharpoonup}{R}_l$ and $\overset{\small\rightharpoonup}{R}_o$ is carried out. The matching method we delegate to the Euclidean distance that has low time-complexity and better performance [36]. Accordingly, the matrix of signal strength disparity (SSD) is.

pyo(19)

In theory, the closer a RP is to the TP, the smaller e value would be. A simple and straightforward way to compute the position of a TP is to search for the minimum e value in E, whose coordinate is assumed to its nearest neighbor. Nevertheless, this pattern does not usually appear due to the ambiguity of the RSSI. Accordingly, we construct the community chain to estimate the position of the tracking tags based on the SSD matrix of E while simultaneously incorporating spatial correlation among the selected RPs. Conceptually, as illustrated in Fig. 2, a community consists of a number of adjoining reference tags, say $2×2$ . The signal strength difference for an arbitrary $m×m$ community is defined as

$c_{i,j}^m = \displaystyle\sum_{x=0}^{m-1}\displaystyle\sum_{y=0}^{m-1}e_{i+x,j+y}$(20)

where m is the rank of the community or the size of the community. And the corresponding matrix cm of SSD for the overall sensing area is expressed as

pyo(21)

Then we introduce a mode for community determination, a hard thresholding. The rank m is flexible, that is, the specific rank can be determined within the range of $λ_{Lo} ≤ H < λ_{Hi}$, where ${λ_{Lo}$ and $λ_{Hi}$ stand for empirical thresholds. In this way, the community can adaptively modulate its rank to raise the probability of enclosing targets in the high interference condition. Therefore, assume a TP is located in the community-m with the minimum value of $c$ in $C$ according to Equation (21). Once the candidate community with rank $m$ is determined, the coordinate ($x,y$) of the TP is computed using

$(x,y) = \displaystyle\sum_{k=1}^{m^2}w_k(x_k,y_k),$(22)

where $(x_k,y_k)$ is the coordinate of the kth RP in the community-m and $w_k$ is the weight to the corresponding tag, which is empirically defined as

pyo(23)

Moreover, we have considered the case with, $e_k = 0$, which has the greatest weight; and others are set to the smallest weight. Doing this ensures that not only the reference tag with the minimum value of e has the greatest weight, but also the estimated coordinate is located within the community-m.

Positioning Algorithm
In this part, we formalize our positioning algorithm that is designed to dynamically accommodate indoor environments for accurate position finding. It comprises six major steps as depicted in Fig. 3.

Fingerprint: This is the process to feature a sensing environment by the proposed active RFID network. The fingerprinting information include the tag’s ID, real coordinates of RP, and corresponding perceived RSSI.

RSSI processing: To alleviate the fluctuation and the ambiguity of RSSI, we introduce a series of processes involving the calibration of RSSI and the noise-canceling of KF.

SSD computation: Once the object is targeted, the system computes the SSD matrix of Equation (19) according to the matching of Euclidean distance.

Rank decision: To figure out the optimal community in terms of rank m, an empirical mechanism is applied according to the RSSI variance of the path loss model.

Community inspection: After the rank of the community is determined, the community-based SSD matrix is computed using Equation (21). An inspection is then performed to acquire the target community with the smallest cost.

Position calculation: Within the enclosing community, the position of the object is calculated by interpolating the surrounding RPs based on Equation (22).



Fig. 3. The diagram of the COMPASS positioning algorithm.


Experimental Results

To thoroughly evaluate our wireless network system for positioning applications, we conducted a series of experiments with simulations and real implementations. Fig. 3 shows the deployment of the scene analysis in a common lab with bookcases, computers and various instruments and working people, where all readers and tags were placed on the lab tables 1.5 m above the ground. There were five TPs randomly distributed amongst an area consisted of $L = 5×5$ RPs. The separation between each pair of RPs was 1 m [31]. Four readers were used in the experiment. They are placed one meter away from the RPs at positions (1,1), (1,J), (I,1) and (I,J).
To quantitatively evaluate the location tracking accuracy, we compute the Euclidean distance between the real and estimated coordinates of the tracking object (tag), i.e.,

$ε=\sqrt{(x-x_r)^2+(y-y_r)^2}$,(24)

where $ε$ is the error of location estimation, $(x,y)$ is the calculated coordinates from Equation (22), and $(x_r,y_r)$ denotes the TP’s true coordinates.
Mean absolute error (MAE) and RMSE are used to analyze the experiment statistics,

$MAE = \frac{1}{n} \displaystyle\sum_{j=1}^{n}|ε_j|$,(25)

$RMSE=\sqrt{\frac{1}{n} \displaystyle\sum_{j=1}^{n}ε_j^2}$,(26)



Fig. 4. A real setup of the reference tags, tracking tags, and RFID readers for the indoor localization experiment.


Simulation
We have conducted a series of simulations based on the environmental deployment of Fig. 4 to validate our proposed positioning algorithm. At first, to realize the impact of maximum time slot $T$ on the real-time PL model, we carried out different environmental conditions with PLE $η$ from 1 to 6 and with an embedded shadowing factor $σ$ from 0 dB to 9 dB. Furthermore, the sliding size $S$ was programmed from 1 to 80 samples with 1 unit increment, i.e., $S =1,2,3,…,80$ samples. Fig. 5 showed that the standard deviation of $H$ reached a stable value after $S = 20$ for every choice $σ$ . We thus chose the slot time $S = 20$ in our proposed system.

As a result of fixing T, we verified the changeover timing of the flexible community rank m. The simulated fingerprint F was generated from the practical PL model with embedded $σ$ from 0 dB to 9 dB. Because H is highly associated with the shadowing factor, we do not emphasize on different PLEs. Fig. 6(a) shows the positioning mean error of various community ranks with respect to different mean $H$ values. Fig. 6(a) and 6(b) reveal that community-2 has a better accuracy if the estimated H is less than $λ_{Lo} = 5.5 dB$ or $H ≤ λ_{Lo}$, whereas community-3 should be chosen if the estimated H is between $λ_Lo = 5.5$ to $λ_Hi = 8 dB$ or $λ_Lo < H ≤ λ_{Hi}$, and community-4 will outperform the other two choices if H is larger than $λ_Hi = 8$ dB or $λ > λ_{Hi}$ .

Fig. 5. Comparison of different sliding size $S$ with varying environmental conditions, including PLE and shadow fading factors. The standard deviation of $H$ reached a stable value after $S = 20$ for every choice $σ$.


Fig. 6. Determination of an optimal community rank. (a) Positioning mean error of various community rank with respect to different PL RMSE values, $H$. (b) The depiction of an optimal rank decision, where the red, blue and purple are the ranges of H that gave better positioning accuracy with the corresponding community ranks.


On-Site Experiments
Fig. 7 shows two circumstances of the environmental dynamics that were recorded from on-site. As illustration, Case 1 and Case 2 show the quantified variance from Equation (10) during a period of time, respectively.

Fig. 7. An illustration of the shadowing effects for different H values. (a) Case 1, $H$ values were below 5.5 dB. (b) Case 2, $H$ values were above 5.5 dB.


To carry out the critical changeover timing on our active RFID system, Fig. 8 elucidates the MAE as a performance index of different community ranks. It can be seen that community-2 indeed offers better accuracy when the H values were below 5.5 dB (Case 1). On the other hand, community-3 outperforms community-2 as the H values increased above 5.5 dB (Case 2). The proposed COMPASS (adaptive community) will automatically choose an optimal community rank that produces a minimal positioning error, i.e., it will select rank 2 for $H ≤ λ_{Lo}$ and rank 3 for $λ_{Lo} < H ≤ λ_{Hi}$, where $λ_{Lo},λ_{Hi}$ can be determined via the results shown in Fig. 6.

Fig. 8. MAE results for applying the changeover timing among different community ranks, community-2, community-3, and adaptive community in realistic experiments. Note, adaptive community chose community-2 for $H ≤ λ_{Lo}$ (case 1), whereas it chose community-3 for a higher $H$ value above 5.5 dB (case 2).


Table 1 lists the positioning error statistics of the three approaches, WkNN, COMPASS, and KF-COMPASS. The 50% error refers to the average error in the prior 50% of ascending order positioning data, with the same explanation for the 75% error. MAE is to evaluate the whole positioning accuracy, and RMSE implies the stability of estimated location of the TP. It can be seen that both COMPASS and KF-COMPASS performed better than traditional WkNN in all categories. The percentage improvements for COMPASS over WkNN were 35.3%, 35.8%, 31.0%, and 28.5%, and KF-COMPASS over WkNN were 49.3%, 46.3%, 37.9%, and 33.9%, respectively. Also, CCA stands for the computational complexity average, which assesses positioning efficiency according to our scenario. The time complexity of COMPASS and traditional WkNN algorithms belong to $O(n^2)$; its positioning time depends on the number of landmarks in the fingerprints, such as density layout. The results showed the COMPASS and KF-COMPASS took 2.7 and 13.5 ms, compared to 2.1 ms of traditional WkNN. Despite an increased computation time, our proposed methodology was still able to perform real-time tracking of multiple objects.

Table 1. Positioning error statistics
Method 50% error (m) 75% error (m) MAE (m) RMSE (m) CCA (ms)
WkNN 1.071 1.405 1.687 1.822 2.1
COMPASS 0.693 0.902 1.164 1.302 2.7
KF-COMPASS 0.543 0.755 1.048 1.204 13.5
Moreover, we compared several recent studies in Table 2. Their positioning frameworks were based on RSSI fingerprinting, which were briefly described in Sections 1 and 2. The network equipment used 5G, BLE and Wi-Fi technologies, respectively. The sensing environments were set within the line of sight. Moreover, the deployed pattern of them all belonged to grid fingerprint and the RP distances were similar to our work. Consequently, our proposed method is practical from the indices of MAE and CDF90 compared to them.

Table 2. Comparison to recent positioning studies based on RSSI fingerprinting
Method 50% error (m) 75% error (m) MAE (m) RMSE (m) CCA (ms)
WkNN 1.071 1.405 1.687 1.822 2.1
COMPASS 0.693 0.902 1.164 1.302 2.7
KF-COMPASS 0.543 0.755 1.048 1.204 13.5
The CDF90 is the estimated error at 90% of cumulative distribution function.


Discussion

Fig. 9 displays the calculated TP positions using WkNN and KF-COMPASS. In Fig. 9(a), the four optimal RPs of red circles are defined as those tags that form the smallest rectangle enclosing the TP. It can be seen from Fig. 9(b), denoted as 0% correctness. When none of the optimal RPs were correctly chosen for the position calculation, the estimated TP positions of WkNN spread over a wide range, and those not enclosed by the optimal RPs would render a large positioning error. Similar situations were found for the 25% correctness and 50% correctness. On the other hand, for the COMPASS approach, the probability of choosing the 0% or 25% optimal RPs were very small. There were about 50% each for correctly choosing 50% and 100% optimal RPs. This is the reason that positioning accuracies are higher compared to the WkNN approach. Note, there was no 75% correctness in the community-2 case because no three out of four RPs can be correctly chosen as shown in Fig. 9(e).
For the WkNN approach, more than two-thirds of the calculations were obtained in the 0% correctness or 25% correctness situations, as shown in Fig. 9(b) and 9(c). And the estimated locations were randomly distributed outside the four optimal RPs. On the other hand, for the KF-COMPASS approach, the odds for correctly choosing the optimal RPs were much higher. About 26% of them occurred in 50% correctness case and 67% happened in 100% correctness case as shown in Fig. 9(d) and 9(f). The estimated locations were relatively stable due to the use of spatial communities instead of individual RP that only depended on RSSI. The results can be found by comparing the MAE and RMSE errors in Table 1.

Fig. 9. Real TP positioning during a certain period of five-minute (300 samples) computed by WkNN and the proposed approach. The colors of black, green, pink, and blue are denoted as RP, TP, estimated locations by KF-COMPASS and WkNN, respectively.


Fig. 10. The probability for the optimal selection of surrounding RPs that closely encompass the TP. 0% correctness means none of the four selected RPs matched the intended optimal RPs, whereas 100% correctness means all of the four selected RPs were exactly the four optimal RPs, i.e., the four tags that form a minimal rectangle enclosing the targeting tag.


Fig. 10 shows the cumulative probabilities of choosing the optimal RPs when calculating the positions of all TPs. As can be seen, for the traditional WkNN approach, it has a high 39.3% chance that none of the four selected RPs were from the optimal RPs, i.e., 0% correctness. And the odds were 37.4% for 25% correctness, 20.2% for 50% correctness and 3.1% for 75% correctness, respectively. Unexpectedly, the correct selection of all four optimal RPs in the WkNN method did not happen. This may be related to the RSSI ambiguity of the experimental environment. On the other hand, 0% correctness probability for COMPASS was only 5.9%, 25% correctness was 3.3%, 50% correctness was 26%, and 100% correctness was as high as 64.8%. Furthermore, the KF-COMPASS improved 0% correctness to 3.7%, 25% correctness to 3.1%, 50% correctness to 26%, and 100% correctness was increased to 67.2%. As a result, our methods promote the correctness of nearest RPs and reduce the influence of ambiguity due to consideration of spatial correlation. Besides, the probability of correctness highly depends on the noise level of the RSSI. As the noise level increases, the probability of correctness for the WkNN approach decreases. Although this is also true for the COMPASS approach, this problem can be readily resolved by choosing a larger community rank, e.g., community-3 versus community-2 to ensure that a higher percentage of correctness can be maintained.


Conclusion

RFID wireless networks have gained a lot of attention for IoT applications. As many of them require accurate positioning, we propose a pragmatic approach, the COMPASS, to improve positioning accuracy. COMPASS divides the sensing area into a number of compact regions called “communities.” The optimal rank of the community can be simply determined via RSSI variance of the adaptive PL model. The intervention of KFs also played a significant part to curb RSSI’s instability, increasing accuracy by 9.3%. Compared to the conventional kNN approach, our experimental results showed that the positioning error using COMPASS (or KF-COMPASS) was about 31%–38% smaller. Furthermore, the statistics in 50% error, 75% error and RMSE were improved by 35%–49%, 36%–46%, and 29%–34%, respectively. On the other hand, we also compared our results to recent positioning studies, and signified that our proposed positioning framework is comparatively practical. The feasibility is due to the promotion of reliability of RSSI and selection of the correct neighborhood.
Because positioning accuracy depends on the severity of environmental influence, further investigations in the large-scale conditions are needed. For example, in order to reduce complexity and computation, a distributed computing approach may be used. The RFID sensor tag with detection of motion and light signals can be used to quickly determine the “residing zone” of the tracking objects and reference points. In addition, RFID magnetic sensors can be fused with our approach to maintain positioning accuracy. In order to deal with larger amounts of data when the network grows increasingly larger, hybrid ring-mesh protocols [40] present a solution to support reliable mesh networks in a large-scale environment. By integrating with wearable sensors, such a framework can be well applied to healthcare applications for accidental fall detection, physiological monitors, and care-needed services.


Author’s Contributions

Conceptualization, YFH, CSC, WCC. Funding acquisition, WCC. Investigation and methodology, YFH, CSC, WCC. Project administration, WCC. Resources, WCC. Supervision, WCC. Writing of the original draft, YFH, CSC. Writing of the review and editing, WCC. Software, YFH, CSC. Validation, YFH, CSC, WCC. Formal analysis, YFH. Data curation, YFH, CSC. Visualization, YFH. All the authors have proofread the final version.


Funding

This work was supported by the Ministry of Science and Technology of Taiwan, ROC (No. MOST 110-2221-E-A49A-505).


Competing Interests

The authors declare that they have no competing interests.


Author Biography

Author
Name : Yung-Fu Hsu
Affiliation : Department of Biomedical Engineering, National YangMing ChiaoTung University, Beitou, Taipei 11221, Taiwan ROC
Biography : received the B.S. and M.S. degrees in biomedical engineering from Chung Yuan Christian University, Taoyuan, Taiwan, in 2007 and 2009. He is currently pursuing the PhD. Degree on biomedical engineering at National YangMing ChiaoTung University, Taipei, Taiwan. His main research interests include signal processing, localization and machine learning. From 2013 to 2017, he was a hardware and software engineer in ATOM Health Corporation in Taiwan. His responsibility includes the development of medical instruments regarding to physiological measurements. During the tenure, he received US patent regarding to cardiovascular monitor device.

Author
Name : Chu-Sung Cheng
Affiliation : Advantech Co., Ltd, Taiwan
Biography : received M.S. degree in biomedical engineering at National YangMing ChiaoTung University, Taipei, Taiwan, in 2009. His research interests include localization, wireless sensor network and ubiquitous computing. Currently, he is a computer engineer in Advantech Co., Ltd, Taiwan.

Author
Name : Woei-Chyn Chu,* Ph.D.
Affiliation : Department of Biomedical Engineering, National YangMing ChiaoTung University, Beitou, Taipei 11221, Taiwan ROC
Biography : Woei-Chyn Chu received the PhD degree in electrical and computer engineering from the University of California at Irvine in 1991. He is distinguished professor in the Department of Biomedical Engineering, National YangMing ChiaoTung University, Taipei, Taiwan, ROC. His major research interests include medical image quantitation/processing, medical device development and medical informatics in healthcare applications. He has published more than 200 research articles and conference presentations in those related fields. He received several patents in integrating RFID technologies in clinical applications and indoor real-time tracking. He also obtained several patents in developing medical devices/methodologies to aid orthopedic surgeries and to assist personal healthcare management. During 2001-2005, he served as the director of the University’s Computer and Communication Center. In 2013-2016 he was the Chairman of the department. He has served as a reviewer, guest editor for a number of international journals and conferences. He is a senior member of the IEEE, member of the ISMRM, Board of Directors of Medical Imaging Standards Association of Taiwan ROC (MISAT), and the Biomedical Engineering Society, Taiwan ROC. He was the President of the MISAT-Taiwan during 2015-2019.


References

[1] F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization systems and technologies,” IEEE Communications Surveys & Tutorials, vol. 21, no. 3, pp. 2568-2599, 2019.
[2] G. M. Mendoza-Silva, J. Torres-Sospedra, and J. Huerta, “A meta-review of indoor positioning systems,” Sensors, vol. 19, no. 20, article no. 4507, 2019. https://doi.org/10.3390/s19204507
[3] J. Kunhoth, A. Karkar, S. Al-Maadeed, and A. Al-Ali, “Indoor positioning and wayfinding systems: a survey,” Human-centric Computing and Information Sciences, vol. 10, article no. 18, 2020. https://doi.org/10.1186/s13673-020-00222-0
[4] K. Ding and P. Jiang, “RFID-based production data analysis in an IoT-enabled smart job-shop,” IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 1, pp. 128-138, 2018.
[5] T. Kamigaki, “Object‐oriented RFID with IoT: a design concept of information systems in manufacturing,” Electronics, vol. 6, no. 1, article no. 14, 2017. https://doi.org/10.3390/electronics6010014
[6] Z. Meng and Z. Li, “RFID tag as a sensor: a review on the innovative designs and applications,” Measurement Science Review, vol. 16, no. 6, pp. 305-315, 2016.
[7] K. Jung and S. Lee, “A systematic review of RFID applications and diffusion: key areas and public policy issues,” Journal of Open Innovation: Technology, Market, and Complexity, vol. 1, no. 1, article no. 9, 2015. https://doi.org/10.1186/s40852-015-0010-z
[8] M. N. K. Boulos and G. Berry, “Real-time locating systems (RTLS) in healthcare: a condensed primer,” International Journal of Health Geographics, vol. 11, article no. 25, 2012. https://doi.org/10.1186/1476-072X-11-25
[9] Y. Alvarez Lopez, J. Franssen, G. Alvarez Narciandi, J. Pagnozzi, I. Gonzalez-Pinto Arrillaga, and F. Las-Heras Andres, “RFID technology for management and tracking: e-health applications,” Sensors, vol. 18, no. 8, article no. 2663, 2018. https://doi.org/10.3390/s18082663
[10] T. K. Geok, K. Z. Aung, M. S. Aung, M. T. Soe, A. Abdaziz, C. P. Liew, F. Hossain, C. P. Tso, and W. H. Yong, “Review of indoor positioning: radio wave technology,” Applied Sciences, vol. 11, no. 1, article no. 279, 2021. https://doi.org/10.3390/app11010279
[11] J. Hightower, R. Want, and G. Borriello, “SpotON: An indoor 3D location sensing technology based on RF signal strength,” University of Washington, Seattle, WA, Technical Report No. 2000-02-02, 2000.
[12] Z. Yang and Y. Liu, “Quality of trilateration: confidence-based iterative localization,” IEEE Transactions on Parallel and Distributed Systems, vol. 21, no. 5, pp. 631-640, 2009.
[13] A. Thaljaoui, T. Val, N. Nasri, and D. Brulin, “BLE localization using RSSI measurements and iRingLA,” in Proceedings of 2015 IEEE International Conference on Industrial Technology (ICIT), Seville, Spain, 2015, pp. 2178-2183.
[14] T. T. Khanh, V. Nguyen, X. Q. Pham, and E. N. Huh, “Wi-Fi indoor positioning and navigation: a cloudlet-based cloud computing approach,” Human-centric Computing and Information Sciences, vol. 10, article no. 32, 2020. https://doi.org/10.1186/s13673-020-00236-8
[15] K. Han and S. H. Cho, “Advanced LANDMARC with adaptive k-nearest algorithm for RFID location system,” in Proceedings of 2010 2nd IEEE International Conference on Network Infrastructure and Digital Content, Beijing, China, 2010, pp. 595-598.
[16] D. Li, B. Zhang, and C. Li, “A feature-scaling-based k-nearest neighbor algorithm for indoor positioning systems,” IEEE Internet of Things Journal, vol. 3, no. 4, pp. 590-597, 2015.
[17] H. Xu, Y. Ding, P. Li, R. Wang, and Y. Li, “An RFID indoor positioning algorithm based on Bayesian probability and K-nearest neighbor,” Sensors, vol. 17, no. 8, article no. 1806, 2017. https://doi.org/10.3390/s17081806
[18] M. B. Afuosi and M. R. Zoghi, “Indoor positioning based on improved weighted KNN for energy management in smart buildings,” Energy and Buildings, vol. 212, article no. 109754, 2020. https://doi.org/10.1016/j.enbuild.2019.109754
[19] M. A. Alsheikh, S. Lin, D. Niyato, and H. P. Tan, “Machine learning in wireless sensor networks: algorithms, strategies, and applications,” IEEE Communications Surveys & Tutorials, vol. 16, no. 4, pp. 1996-2018, 2014.
[20] S. Bozkurt, G. Elibol, S. Gunal, and U. Yayan, “A comparative study on machine learning algorithms for indoor positioning,” in Proceedings of 2015 International Symposium on Innovations in Intelligent SysTems and Applications (INISTA), Madrid, Spain, 2015, pp. 1-8.
[21] S. Lee, J. Kim, and N. Moon, “Random forest and WiFi fingerprint-based indoor location recognition system using smart watch,” Human-centric Computing and Information Sciences, vol. 9, article no. 6, 2019. https://doi.org/10.1186/s13673-019-0168-7
[22] E. Homayounvala, M. Nabati, R. Shahbazian, S. A. Ghorashi, and V. Moghtadaiee, “A novel smartphone application for indoor positioning of users based on machine learning,” in Adjunct Proceedings of the 2019 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2019 ACM International Symposium on Wearable Computers, London, UK, 2019, pp. 430-437.
[23] G. Li, E. Geng, Z. Ye, Y. Xu, J. Lin, and Y. Pang, “Indoor positioning algorithm based on the improved RSSI distance model,” Sensors, vol. 18, no. 9, article no. 2820, 2018. https://doi.org/10.3390/s18092820
[24] G. Felix, M. Siller, and E. N. Alvarez, “A fingerprinting indoor localization algorithm based deep learning,” in Proceedings of 2016 8th International Conference on Ubiquitous and Future Networks (ICUFN), Vienna, Austria, 2016, pp. 1006-1011.
[25] M. T. Hoang, B. Yuen, X. Dong, T. Lu, R. Westendorp, and K. Reddy, “Recurrent neural networks for accurate RSSI indoor localization,” IEEE Internet of Things Journal, vol. 6, no. 6, pp. 10639-10651, 2019.
[26] A. Riady and G. P. Kusuma, “Indoor positioning system using hybrid method of fingerprinting and pedestrian dead reckoning,” Journal of King Saud University-Computer and Information Sciences, 2021. https://doi.org/10.1016/j.jksuci.2021.09.005
[27] T. Rappaport, Wireless Communications: Principles and Practice, 2nd ed. New Jersey, NJ: Prentice-Hall, 2001.
[28] A. Mackey, P. Spachos, L. Song, and K. K. Plataniotis, “Improving BLE beacon proximity estimation accuracy through Bayesian filtering,” IEEE Internet of Things Journal, vol. 7, no. 4, pp. 3160-3169, 2020.
[29] C. H. Huang, L. H. Lee, C. C. Ho, L. L. Wu, and Z. H. Lai, “Real-time RFID indoor positioning system based on Kalman-filter drift removal and Heron-bilateration location estimation,” IEEE Transactions on Instrumentation and Measurement, vol. 64, no. 3, pp. 728-739, 2015.
[30] L. M. Ni, Y. Liu, Y. C. Lau, and A. P. Patil, “LANDMARC: indoor location sensing using active RFID,” in Proceedings of the 1st IEEE International Conference on Pervasive Computing and Communications, Fort Worth, TX, 2003, pp. 407-415.
[31] P. W. Hsu, T. H. Lin, H. H. Chang, Y. T. Chen, Y. J. Tseng, C. H. Hsiao, et al, “Simulations and experiments for optimal deployment of an RFID‐based location‐aware system,” Wireless Communications and Mobile Computing, vol. 11, no. 6, pp. 679-691, 2001.
[32] A. Ye, Y. Ling, L. Xu, and X. Yang, “An improved RFID-based localization algorithm for wireless sensor networks,” International Journal of Distributed Sensor Networks, vol. 2013, article no. 390194, 2013. https://doi.org/10.1155/2013/390194
[33] Y. Luo, O. Hoeber, and Y. Chen, “Enhancing Wi-Fi fingerprinting for indoor positioning using human-centric collaborative feedback,” Human-centric Computing and Information Sciences, vol. 3, article no. 2, 2013. https://doi.org/10.1186/2192-1962-3-2
[34] B. Wang, X. Liu, B. Yu, R. Jia, and X. Gan, “An improved WiFi positioning method based on fingerprint clustering and signal weighted Euclidean distance,” Sensors, vol. 19, no. 10, article no. 2300, 2019. https://doi.org/10.3390/s19102300
[35] Y. Xie, Y. Wang, A. Nallanathan, and L. Wang, “An improved K-nearest-neighbor indoor localization method based on spearman distance,” IEEE Signal Processing Letters, vol. 23, no. 3, pp. 351-355, 2016.
[36] J. Li, X. Gao, Z. Hu, H. Wang, T. Cao, and L. Yu, “Indoor localization method based on regional division with IFCM,” Electronics, vol. 8, no. 5, article no. 559, 2019. https://doi.org/10.3390/electronics8050559
[37] S. Huang, K. Zhao, Z. Zheng, W. Ji, T. Li, and X. Liao, “An optimized fingerprinting-based indoor positioning with Kalman filter and universal kriging for 5G Internet of Things,” Wireless Communications and Mobile Computing, vol. 2021, article no. 9936706, 2021. https://doi.org/10.1155/2021/9936706
[38] J. Wang and J. Park, “An enhanced indoor positioning algorithm based on fingerprint using fine-grained CSI and RSSI measurements of IEEE 802.11n WLAN,” Sensors, vol. 21, no. 8, article no. 2769, 2021. https://doi.org/10.3390/s21082769
[39] L. Garg, E. Chukwu, N. Nasser, C. Chakraborty, and G. Garg, “Anonymity preserving IoT-based COVID-19 and other infectious disease contact tracing model,” IEEE Access, vol. 8, pp. 159402-159414, 2020.
[40] H. H. Attar, A. A. Solyman, A. Alrosan, C. Chakraborty, and M. R. Khosravi, “Deterministic cooperative hybrid ring-mesh network coding for big data transmission over lossy channels in 5G networks,” EURASIP Journal on Wireless Communications and Networking, vol. 2021, article no. 159, 2021. https://doi.org/10.1186/s13638-021-02032-z

About this article
Cite this article

Yung-Fu Hsu, Chu-Sung Cheng, and Woei-Chyn Chu*, COMPASS: An Active RFID-Based Real-Time Indoor Positioning System, Article number: 12:07 (2022) Cite this article 1 Accesses

Download citation
  • Recived8 August 2021
  • Accepted23 December 2021
  • Published15 February 2022
Share this article

Anyone you share the following link with will be able to read this content:

Provided by the Springer Nature SharedIt content-sharing initiative

Keywords