Advertisement

Deployment of RSS-Based Indoor Positioning Systems

  • Christian Esposito
  • Massimo Ficco
Article

Abstract

Location estimation based on Received Signal Strength (RSS) is the prevalent method in indoor positioning. For such positioning systems, a massive collection of training samples is needed for their calibration. The accuracy of these methods is directly related to the placement of the reference points and the radio map used to compute the device location. Traditionally, deploying the reference points and building the radio map require human intervention and are extremely time-consuming. In this paper we present an approach to reduce the manual calibration efforts needed to deploy an RSS-based localization system, both when using only one RF technology or when using a combination of RF technologies. It is an automatic approach both to build a radio map in a given workspace by means of a signal propagation model, and to assess the system calibration that best fits the required accuracy by using a multi-objective genetic algorithm.

Keywords

Positioning systems Propagation models Optimization solvers 

1 Introduction

The recent progress in portable mobile devices and the proliferation of pervasive applications made context-aware systems of growing interest. Context-awareness is related to the ability of a system to adapt content and presentation of provided services to the user context [1]. Such a context can be defined by means of several elements, but location represents the most considered one. So, in the last years research has spent a massive amount of efforts to devise location-sensing technologies and to support location-aware applications [2]. The key problem to provide location-awareness within the context of mobile computing is to compute the exact physical location of the user; such a problem is named positioning. If positioning in outdoor scenarios is no more an issue thanks to widely adopted solutions, when it comes to deal with indoor scenarios, things have not been settled down and positioning in such environments is still considered an open issue.

In the last years, several positioning systems for indoor environments have been proposed [3], and they can be characterized in terms of infrastructure costs, i.e., efforts made to design and setup the infrastructure in a given workspace, and accuracy, i.e., how well the system is able to estimate the user location. As illustrated in Fig. 1, a positioning system performs a partitioning of the workspace in a certain number of areas, where all the points share the same localization information that is equal to the location at the center of the area. Therefore, indoor positioning systems aim at assigning an area identifier as user location, implying a localization error equal to the distance between the real user location and the center of the assigned area. So, the margin of localization error for a given area can be calculated as the distance between its points on the border and its center. Accuracy consists of the maximum margin along all the areas within the workspace, and is in inverse proportion to the positioning quality provided by a certain system (higher is the accuracy, worse is the quality of the positioning system).
Fig. 1

Localization error for a generic positioning system

Indoor positioning systems can be classified in two distinct groups, which differ with respect to offered accuracy and required infrastructure costs. The first group contains those positioning systems that adopt dedicated infrastructures (e.g., based on ultrasound emissions or ultra-wide band), whose only scope is to support positioning operations. On the other hand, the second group contains those positioning systems that adopt non-dedicated technologies (e.g., Bluetooth, WiFi or RFID), whose primary function is not related to positioning, but they have been deployed for communication purposes. Using a dedicated infrastructure allows exposing lower accuracy, but also exhibiting strong costs since an ex-novo infrastructure needs to be deployed. On the other hand, the spreading of wireless hot-spots into many public and private places, as well as the new generation of mobile devices supporting several wireless technologies (e.g., the Nokia C7-00 1 supports both WiFi and Bluetooth, or new Mobile RFID Reader DL710 2 also provides wireless transfer functions by means of WiFi and Bluetooth) has fostered the development of indoor positioning systems based on standard wireless communication technologies [4]. In fact, both academia and industry have developed several indoor positioning systems based on non-dedicated infrastructures [5], which mostly adopt Received Signal Strength (RSS) measures of Radio Frequency (RF) signals as a mean to compute the user location. However, such systems are affected by some serious issues on calibration efforts and achievable accuracy. Specifically, a study presented in [6] has proved that such limits in the accuracy offered by RSS-based positioning systems are intrinsic due to the use of standard RF technologies and is unlike to be further improved if more complex approaches and additional infrastructure are not introduced. To reduce the accuracy exhibited by RSS-based positioning solutions, hybrid approaches have been proposed [7]. They compute user location by combining location information obtained by Access Points (APs), namely reference points or sensors, of different wireless technologies.

A generic RSS-based positioning system usually adopts a Fingerprinting localization technique, which performs the following operations: measuring the strength of the signals propagated from a set of reference points, and using the collected measures to compute user position. Such a system is characterized by two distinct operational phases [4]. During an off-line phase, the system is calibrated by deploying the given reference points in the workspace, and building the so-called Radio Map (RM), i.e., a table mapping RSS samples measured at a number of known locations to the respective locations. During an on-line phase, strengths of the signals generated from the reference points are measured by the user device, then they are used to compute device/user location by using the built RM. The user location is assumed equal to the position in the RM related to the RSS values that are numerically closest to the measured ones. Specifically, a RSS sample has the form of a vector, such as \(\langle val_{0}, val_{1}, \ldots, val_{n}\rangle\), when n sensors are deployed and where \(val_{i}\) is the average of RSS values of signals received from the i-th sensor. Then, the nearness between two RSS samples, i.e., one obtained from massive measurements, namely \(RSS\), and one retrieved from the RM, namely \(rss\), can be computed as the euclidian distance between two vectors:
$$ d = \sqrt{(RSS_{0} - rss_{0}) + \cdots + (RSS_{n} - rss_{n})} $$
(1)
The sample in the RM with the lowest value for Eq. 1 with respect to \(RSS\) is assumed as the numerically closest one, and its associated location information is assigned to the user.

The accuracy of an indoor positioning system is affected by how well the system has been configured, i.e., how the RM has been constructed and where reference points are placed. Since the deployment of these APs is not usually made by considering positioning demands, but considering the signal coverage of the overall workspace, the accuracy of the location system may be poor. Using more than one wireless technology allows improving accuracy, but, on the other hand, it further increases the complexity of the deployment and the calibrating of indoor location-sensing infrastructures. Specifically, reference points of different technologies have to be placed in order to maximize the accuracy of positioning as well as to minimize any interference phenomena that may arise among them. Moreover, the number of reference points per each technology has to be chosen in order to maximize the achievable accuracy and to reduce the infrastructure costs. To the best of our knowledge, we argue that such problem has not been fully investigated by the research community. A paper [8] has been recently published on such issues, but it still requires a massive usage of measurements and deals only with the RM construction, while it does not treat sensor placement. An interesting industrial product is Ekahau Real Time Location System (RTLS)3, which is fully integrated with Ekahau Site Survey (ESS)4 to minimize deployment time and cost. However, it is tailored only on 802.11a/b/g/n Wi-Fi networks, it does not consider the possibility of using more than one RF technology and it is not equipped with an automatic approach for choosing the best deployment.

In this work, we present an approach aiming at reducing the efforts spent during the calibration phase, and at improving the overall accuracy of the system. In particular, the contribution of this paper is twofold:
  1. 1.

    limiting the human intervention when building the RM by modeling the signal propagation in the workspace, and reducing the need of a massive campaign to collect RSS samples;

     
  2. 2.
    proposing a methodology to assist the system deployer in the choice of the best placement of the reference points. In particular, possible scenarios that we can consider are the following ones:
    • The system makes use of infrastructures deployed only for positioning demands. The issue is to estimate how many and what kind of wireless reference points are needed, and where they have to be placed in order to achieve an accuracy that fits the user requirements.

    • The adopted infrastructure consists of classical wireless APs already deployed in the workspace for communications purposes. The issue is yet to define how many additional, what kind of reference points are needed, and where to place them, so to improve system accuracy.

     

We have proposed to formulate the issue of deploying a positioning system as an optimization problem. Specifically, we adopt an approach based on a genetic algorithm in order to automatically select the best configuration (i.e., number, kind and locations of the reference points to place in the workspace) that fits the user/application requirements in terms of accuracy. In addition, we have used an analytical model of the signal propagation in order to reduce off-line manual efforts needed to build an RM.

The rest of the paper is organized as follows: Section 2 provides a detailed survey of positioning approaches in indoor environments, examining the most adopted ones and the reasons of their success. Section 3 illustrates the issues of calibrating an RSS-based indoor positioning system. It is composed of two subsections: one is about the problems related to calibration, while the other one discusses the state of the art of solutions to the problems formulated in the first subsection. Section 4 presents our approach and how it can be used to built RMs and/or to select the optimal placement pattern. Section 5 speaks about how signal propagation is modeled so to reduce the need of on-field measurements, while Section 6 explains how placement pattern is selected in an automatic manner. Section 7 discusses experimental campaigns we have conducted to assess the effectiveness of our approach. Last, we conclude with Section 8, which contains final remarks on our approach and future work.

2 Indoor Positioning System

The most recognized positioning system uses signals from satellites to estimate user location (satellite positioning) since nowadays almost every smartphone is equipped with a GPS receiver [9]. However, satellite signals exhibit poor coverage in indoor environments [4]. The other widely known positioning solution uses the GSM/CDMA mobile cellular network to estimate user location with respect to the cell-ID (cellular positioning). However, even such a solution poorly performs [3]. In fact, their accuracy depends on the cell size, which is within the range of 50-200 meters. As it is intuitive, methods commonly used in outdoor environments face scarse applicability or too high accuracy in indoor environments, so they have been poorly considered to address the issue of estimating the user location in indoor scenarios, as shown in Fig. 2a. In fact, only 4% of all the available systems on indoor positioning make use of such methods.
Fig. 2

Study of the literature on indoor positioning considering around 70 papers collected from [3, 4, 10]: (a) taxonomy of the available papers with respect to the adopted technologies for location estimation; (b) summary of the accuracy of the main current indoor positioning systems, which has been computed as the mean of the accuracy exhibited by the systems surveyed in Table 1; and (c) adoption of certain positioning techniques with given RF wireless technologies

During the recent years, there have been several research projects and industrial solutions that aimed at designing better solutions for the indoor positioning [3, 4, 5]. The first kind of such solutions makes use of infrastructures specifically deployed for positioning purposes. Several practical examples of such kind of systems can be found in the literature, and are mainly based on the following technologies:
  • InfraRed (IR) Positioning [11]: Infrared technology is adopted to perform location sensing by using short-range narrow-transmission-angle beam.

  • UltraSound (US) Positioning [12]: ultrasonic technology is used to measure the location of a tag carried by the user, similarly to the navigating system adopted by bats to move in the darkness. Specifically, a set of receivers are mounted on the ceiling of the indoor environment at known locations, and the user tag periodically broadcasts a short pulse of ultrasound signals.

  • Ultra Wide Band (UWB) [13]: it requires sending ultrashort pulses (i.e., with a bandwidth lower than 500 MHz).

  • Vision-based Positioning [14]: several cameras are used to analyze the scene and figure out where the user is located.

  • Magnetic-based Positioning [15]: DC magnetic fields are adopted to locate an user device.

As clearly illustrated in Fig. 2b and Table 1, this first class of positioning systems solutions is extremely accurate, i.e., they provide the lowest accuracy that is between 0.1mm and around 1m. However, Fig. 2a reveals that they are not widely adopted (in fact, only about 26% of the solutions available on indoor positioning belong to this first class), while the largest part (e.g., about 70%) of the positioning systems devised by research and commercialized by industries uses standard wireless networking technologies, i.e., Bluetooth (BT), WiFi, RFID or Ultra High Frequency (UHF). In fact, such systems with ad-hoc infrastructures grant low accuracy at the expenses of a too strong financial investment. Therefore, it is more convenient the use of technologies that perform other duties jointly to positioning ones. There have been many attempts to realize RF-based positioning systems based on Bluetooth (BT), WiFi, RFID or Ultra High Frequency (UHF). Figure 2a illustrates a possible classification of such systems. Despite the systems based on RF technologies are considered promising solutions for indoor environments, they could not meet the same success as GPS in outdoor environment due to a severe drawback: as illustrated in Fig. 2b, they exhibit a higher accuracy than ad-hoc solutions, i.e., achievable accuracy fluctuates between 1.5 m and 3.76 m with BT, between 2 m and 3 m with RFID, and from 1 m to 5,4 m with WiFi.
Table 1

Survey of indoor positioning systems

Prototype

Technology

Accuracy

Active Badge [3]

IR

n.a.

Firefly [3]

IR

3.0 mm

OPTOTRACK PROseries [3]

IR

0.1–0.5 mm

IRIS_LPS [3]

IR

16 cm

Active Bat [3]

US

3 cm

Cricket [3]

US

10 cm

Sonitor [3]

US

n.a.

Ubisense [3]

UWB

15 cm

RADAR [3]

WIFI

3–5 m

Horus [4]

WIFI

2 m

DIT [4]

WIFI

3 m

PinPoint 3D-iD [4]

WIFI

1–3 m

Ekahau [4]

WiFi

1 m

Robot-based [4]

WIFI

1.5 m

MultiLoc [4]

WIFI

2.7 m

TIX [4]

WIFI

5.4 m

COMPAS [3]

WIFI

1.65 m

Topaz [3]

BT

2–3 m

 [16]

BT

2.06 m

 [17]

BT

1.5 m

BLPA

BT

3.76 m

WhereNet [3]

RFID

2–3 m

LANDMARC [4]

RFID

<2 m

SpotOn [4]

RFID

n.a.

SmartLOCUS [4]

US + RF

2–15 cm

EIRIS [4]

IR + RF

<1 m

 [18]

RFID + WIFI

260–554 cm

 [19]

RFID + BT

0.25–2.25 m

 [20]

BT + WIFI

2–2.5 m

A solution to improve the high accuracy of RF-based systems is to combine heterogeneous technologies for the location estimation (Hybrid Positioning) [7]. Such systems have shown a sensible improvement in the accuracy compared to using only one single RF technology (e.g., an accuracy of 2-2.5 meters with the combined use of BT and WiFi, 260-554 centimeters with WiFi and RFID, and 0,25-2,25 meters with BT and RFID, as listed in Table 1). Using more than one RF technology is not only motivated by accuracy concerns. In fact, a workspace may not be uniformly covered by the RF signals of a given technology, so a positioning system is required to be opportunistic by using the specific technologies that are available in the area within which the user is moving.

A generic positioning system is architected according to a layered structure [21], as depicted in Fig. 3, where technologies stand at the bottom. On top of them there are the so-called methods, which represent the features measured by using the adopted technologies. The following ones are the most cited methods in the literature [5]:
  • Time of Arrival (ToA): time taken by a signal to reach its destination;

  • Time Difference of Arrival (TDoA): difference among reception instances of a signal at two, or more, spatially separated receiver devices;

  • Angle of Arrival (AoA): estimation of the direction from which a device receives a signal emitted by a given source;

  • Received Signal Strength (RSS): power of signals received by a given device.

Fig. 3

Layered structure of a generic positioning system

As illustrated in Fig. 3, on top of the methods we find the techniques, which specify how to use raw measures to compute the user location. In the literature, there have been proposed several different techniques; however, the most known, and widely adopted, are the following ones [5]:
  • Triangulation, or multilateration: the user location is computed from given measures of angles and sides of one or more triangles formed by the point where the user is located and two, or more, so-called anchors, i.e., sensors at known location;

  • Fingerprinting: the strength of the signals received from some reference points is sampled in some areas of the workspace with known location; such samples are called fingerprints. The user position is equal to the area with fingerprint closer to the signal strength measured in the location where the user is placed. This simple approach has been further evolved by more sophisticated ones. A concrete example is the so called Probabilistic Fingerprinting Approach, such as [22]: the location information of each fingerprint is summarized by Probability Density Functions, so to create a probability model of the workspace. Pattern Recognition techniques are used to infer the user location from the probability model by using measured RSS values;

  • Proximity: the user position is equal to the one of the nearest emitter with known location, and a nearness measure can be formulated in terms of \((i)\) signal time of arrival, i.e., if an emitter is closer to the user device than the other ones, its signals will be the first ones to be received; and \((ii)\) signal reachability, i.e., the user device will receive signals from a given emitter only if located nearby due to the short range of the adopted technology. In case more than one sensor is detected as the closest to the user device (e.g., it is reached by beacons of more than one sensor with the same ToA), then Triangulation can be used to resolve this situation.

Not all the possible combinations between technologies, methods and techniques are feasible in real indoor positioning systems. For a concrete example, due to the short range of the infrared radiation, IR-based systems mainly make use of proximity to estimate user location, or the distance between the ultra-sound tags and receivers can be computed through the ToA of the pulse, from which user location is obtained by using the ToA information of at least three receivers by using simple geometrical rules that go back to the Pitagora theorem. With respect to RF-based systems, we have performed an analysis of the most adopted techniques. As shown in Fig. 2c, such positioning systems mostly use Fingerprinting as technique and RSS as method. For this reason, this paper is focused on them; however, any other technique related to RSS is also applicable.

3 Deployment and Configuration Concerns for Indoor RSS-Based Positioning

As mentioned, the quality of a given positioning system is formulated in terms of accuracy, which indicates how close the estimated locations are to the real one, and is strongly affected by how well it has been configured, i.e., how adopted RM has been constructed, and where reference points have been placed. In addition, since more than one wireless technology may be used to improve accuracy, the complexity of deployment and calibration of indoor location-sensing infrastructures is further increased. Specifically, reference points of different technologies have to be placed so maximize the achievable accuracy as well as to minimize the infrastructure costs.

3.1 Problem Statement

RSS-based indoor positioning systems are characterized by two distinct operational phases [4]: an off-line phase, where the system is configured by building RM; and an on-line phase, where RSS samples are measured and used to compute user location by using the built RM. Therefore, in the off-line phase, we have to deal with two distinct problems so to properly calibrate the given indoor positioning system:
  • RM Construction: given a set of reference points, characterized by the same wireless technology or not, we have to obtain a characterization of the signal propagation within the workspace so to extract RSS patterns at known locations and build an RM;

  • Optimal Placement Pattern: we have to decide which technologies to use, how many reference points per technology to deploy, and where such sensors need to be located so to achieve the optimal positioning accuracy.

Both problems strongly require a computer-aided approach in order to reduce the effort of manually tuning the system by automatically generating solutions that best fit the user requirements in terms of accuracy by limiting the usage of on-field measurements. The following two paragraphs provide a detailed description of these two problems.

3.1.1 RM Construction

The solution to the first issue traditionally begins by dividing the workspace in several areas, whose number is indicate with \(N_{A}\). Then, since signal strength varies noticeably due to interferences and environment conditions, several measurements need to be collected in a single point (i.e., at the center of each area), and we indicate with \(N_{S}\) the number of samples collected in a given location. If we indicate with \(T_{M}\) the time needed to measure an RSS sample from all the reference points deployed within the workspace, the time to calibrate the system, namely \(T_{C}\), is formulated as follows [8]:
$$ T_{C} = N_{A} \times N_{S} \times T_{M} + \epsilon, $$
(2)
where \(\epsilon\) is not negligible (it is even possible that it is higher than the time spent to record RSS values, i.e., the first addend in the equation) and represents the time spent for moving between the calibration points and/or performing other tasks during which the system is not actively recording RSS values.

Traditional measurements-based solutions to build RMs present a pivotal requirement that can be expressed by the following question: “is it possible to reduce to the minimum the number of needed measurements without compromising the accuracy of the obtained RM?”. In fact, conducting an heavy measurement campaign requires strong efforts in terms of time spent to perform it and money spent to have a person measuring RSS patterns. When tuning the system, a requirement is to keep lower the needed measurements so to minimize the quantity \(T_{C}\) in Eq. 2, and to restrain calibration efforts.

3.1.2 Placement Pattern Selection

When a newly RSS-based indoor positioning system has to be deployed, it is crucial to define how many sensors have to be used and where they have to be located (if the system is hybrid such choice must be made per each used technology). Moreover, such issue raises not only at the beginning of the positioning system life cycle, but also in other cases:
  1. 1.

    one or more sensors may fail due to hardware malfunctioning, so other ones need to be placed in order to prevent a drop in accuracy;

     
  2. 2.

    one or more sensors have to be moved to new locations without compromising the overall accuracy of the positioning system;

     
  3. 3.

    one or more sensors need to be included in a proper manner so to improve the positioning accuracy.

     

Selection of the optimal placement pattern exhibits two key requirements. On one hand, it is important to achieve a reasonable radio coverage of the workspace. If there is no coverage in a particular area of the workspace, the system may not be able to compute the user position in there. On the other hand, the system tuned with the selected pattern has to expose an optimal positioning accuracy. Unfortunately, no general guidelines exist when configuring these systems. In fact, the goodness of a placement pattern highly depends on the specific workspace conditions, i.e., wall position and material, space topography, noise sources, number and disposition of the people that attend the place.

For identifying an optimal placement pattern, the positioning system installer or designer should typically perform a massive field measurement campaign. In particular, the installer has to perform the following operations per each chosen placement patter:
  1. 1.

    arrange the sensors according to the given pattern;

     
  2. 2.

    measure wireless coverage and accuracy in a set of given points within the workspace;

     
  3. 3.

    assign to the placement pattern a grade that expresses its goodness (which reflects the accuracy of the positioning system deployed by using the given placement pattern).

     
After examining an enough-large set of placement patterns, the installer can determine the optimal as the one with the best grade and use it to deploy the system. The total time spent to select the optimal placement, namely \(T_{D}\), is equal to the following equation, where \(T_{C}\) is defined in Eq. 2:
$$ T_{D} = \nu \times (T_{C} + \delta) \quad with \, \nu \ll (N_{AP})^{N_{A}}, $$
(3)
where \(\nu\) is the number of examined patterns, whose maximum number, namely \(\nu_{max}\), is equal to \((N_{AP})^{N_{A}}\) (where \(N_{AP}\) is the maximum number of sensors deployable within the working area) but usually is equal to an integer that is much lower than \(\nu_{max}\). In the previous equation \(\delta\) is not a negligible contribution to \(T_{D}\), and represents the time needed to deploy the sensors given a certain placement pattern (i.e., putting in place the required network infrastructures, such as electricity wires or networking cables, and hanging the sensors on walls and ceilings). This method is evident to be costly, time-consuming and error-prone. In fact, efficiency (i.e., how fast the selection is performed) is in conflict with effectiveness (i.e., how good the selection is performed): when \(\nu\) is higher, the approach is less efficient (since \(T_{D}\) increases), but also more effective (since it is probable to find better solutions than the cases with a lower \(\nu\) due to the possibility of evaluating more solutions). The trade-off among the two characteristics is a very challenging issue that does not have a solution without the support of an automatic tool to find the most suitable value for \(\nu\) (i.e., higher values can not further improve the effectiveness over a given threshold while keeping \(T_{D}\) within acceptable value).

3.2 Available Solutions for Deploying and Tuning Indoor RSS-Based Positioning Systems

In this section, we focus on presenting the state of the art on approaches to deal with the previous problems.

3.2.1 Approaches for RM Construction

Since \(T_{M}\) in the Eq. 2 only depends on the adopted technologies, and \(\epsilon\) on the size of the working area and \(N_{A}\), the traditional ways to lower the calibration efforts are (1) reducing the number of the areas in which the workspace is partitioned (reduction of \(N_{A}\)), and/or (2) spending less time to collect RSS samples at each area (reduction of \(N_{S}\)) [23]. However, such solutions may bring to RMs of lower quality, i.e., it does not fully characterize the features of signal propagation within the workspace, with the consequence to compromise the quality of the overall positioning system [8].

An other solution is to predict RSS values values by means of modeling signal propagation. Ekahau Site Survey represents a professional tool to simulate signal propagation features given a set of sensors deployed within a certain workspace. This tool has been mainly designed for ground-level view of coverage and performance of Wi-Fi networks, without considering any positioning concerns. Also other works, such as [24], propose to use Ray Tracing [25] as a mean to model signal propagation. It considers an AP as launching rays, each is considered locally straight for a certain, even small, distance, (e.g., until a solid obstacle is found), after which the new direction of the ray is calculated, and a new ray is sent out. As the simulation advances, the strength of the ray is altered. The process is repeated until a complete path is generated and all the workspace is covered. The weaknesses of this solution are the following ones:
  1. 1.

    They are tailored on a given RF technology and do not consider the case where more than one RF technology is used within the positioning system;

     
  2. 2.

    Ray Tracing is an accurate method for indoor signal strength estimation; however, its overhead is quite high with respect to other estimation methods [26].

     

Since we consider RM construction as a building block of the overall approach to efficient calibration of a given positioning system, we seek a simple and fast method for modeling signal propagation even if this choice slightly compromises the achievable accuracy.

3.2.2 Approaches for Placement Pattern Selection

As formulated in Eq. 3, the calibration time, i.e., \(T_{D}\), is function of the number of solutions to be evaluated, i.e., \(\nu\), and the time to construct an RM for a given placement pattern, i.e., \(T_{C}\). The possible ways to reduce the calibration time is, therefore, to reduce \(\nu\) and/or \(T_{C}\).

A first solution, such as Ekahau Site Survey, consists of introducing a model-based approach to construct the RM and minimizing the need of an on-field measurement campaign. Despite being successful to address the issue of deploying APs for communication concerns, such a solution presents three drawbacks if we want to use it for dealing with the placement pattern selection: \((i)\) arrangements are made depending operator judgement and not via an automatic approach (i.e., the operator has to select a certain number of possible deployments and evaluate it with the tool), and \((ii)\) only signal coverage is considered, which is not the only factor that negatively affects positioning accuracy, and \((iii)\) only WiFi is considered.

An other solution is to lower \(\nu\) without compromising the effectiveness of the placement selection. However, to the best of our knowledge, there are no works on positioning calibration that adopt such solution. On the contrary, there are other works in related fields that apply such approach. Specifically, deployment issues in RSS-based indoor positioning systems is closely related to other well-known planning problems:
  • Planning Process for Cellular Networks [27], which consists of two sub-problems: Antenna Positioning Problem (APP), i.e., deciding the site location for the antennas, number or type of antennas per each site, and other antenna parameters; and Frequency Assignment Problem, i.e., selecting which available frequencies to assign to the antennas in the network.

  • Design Wireless ad-hoc Networks in Indoor Environments [28], whose main objective is to find a configuration of APs so to assure high coverage, low interference level or minimum available throughput.

  • Coverage Problems in Wireless Sensor Networks [29], which aims at deploying sensors in a way that they are able to observe a certain physical space in an appropriate manner, i.e., at least one sensor covers each location of the physical space of interest. Moreover, deployment is performed by considering additional requirements rather than only coverage, e.g., sensor sensing range may be dynamically adjusted so to conserve energy resources without compromising the sensing coverage objective.

Such problems have been extensively studied in the last decades by formulating them in terms of optimization problems and resolving them by means of highly effective optimization algorithms, such as exact solving approaches or heuristic ones. The fundamental question we have asked to ourselves is “Could we apply approaches contained in such a rich literature to address the calibration issues in RSS-based indoor positioning systems?”. The only answer we came up is “No". In fact, as proved by [30], a placement pattern chosen for only communication purposes (i.e., by maximizing signal coverage or provided QoS), without taking care of positioning concerns, can not be optimal for positioning systems. In fact, such paper shows that even the best placement for communication purposes exhibited a non appropriate accuracy, e.g., around 8 meters, which is completely not suitable for indoor positioning. Our approach to deal with calibration issues in RSS-based indoor positioning systems is to drawn on the experience of planning approaches for wireless and cellular networks and to extend them by including positioning concerns in the formulation of the overall optimization problem and by inserting positioning model-based emulation in the adopted solving approach.

A last solution is to do not perform the selection process, and to select sensor deployment according to specific guidelines. Literature on US-based positioning contains some papers, such as [31], which provide guidelines to where locate sensors, but they are not applicable for indoor RF-based systems, due to the strong differences in terms of propagation features between RF and UD signals. Moreover, an automatic approach to deploy US-based sensors is also described in [32]. Despite the fact of being tailored on US-based positioning systems and considering TOA technique instead of Fingerprinting, this approach does not fit to our case since it aims to optimize only signal coverage, which, as we have previously said, is not a winning choice for RSS-based indoor positioning systems.

4 An Automated Strategy for Deploying RSS-Based Positioning Systems

This section describes the approach that we propose to deal with the issue of calibrating RSS-based positioning systems. Figure 4 provides a general overview of our approach. Specifically, it consists of three components: the first one, called Workspace Characterizer, has the scope of dialoging with the user and obtaining a description of the workspace, adopted sensors and chosen positioning technique; the second one, called RM Builder, constructs RMs starting from user inputs, and can be also teamed up with a third one, named Placement Pattern Finder, that handles the selection of the best deployment configuration.
Fig. 4

Overview of the proposed approach

The rest of this section is structured in two subsections. Section 4.1 describes how the proposed approach is used to build RMs without wasting time to perform any heavy measurement campaigns. Whereas, Section 4.2 illustrates how our approach is used to optimally place reference points in the workspace without requiring an inefficient traditional way of guessing the best deployment by trial and error.

4.1 Building Radio Maps by Reducing On-field Measurements

As illustrated in Fig. 4, RM Builder is composed of two elements: an RSS Estimator, which returns a characterization of RSS patterns within the workspace given a certain placement pattern, and RM Synthesizer, which constructs an RM starting from the RSS patterns provided as input. In the traditional approaches, the RSS Estimator consists of massive on-field measurement campaigns, while the RM Synthesizer implements the adopted positioning technique, such as Fingerprinting, with one or more technologies. To avoid the recourse to time-consuming measurements, we propose to implement in the RSS Estimator a model of signal propagation features in the workspace.

Given a model for the RF signal propagation, RM Builder executes our approach for building an RM with performing the following steps, as shown in Fig. 5:
  • Step-1: the user provides a description of the work-space, in terms of the following factors:
    • space topography, i.e., shape and size of the work-space, and location and type of walls and doors;

    • deployment, i.e., number and position of the reference points;

    • tessellation, i.e., size of the areas in which the workspace is partitioned.

  • Step-2: the graphical characterization of the work-space is transformed into a mathematical matrix, which can be used to resolve an analytical signal propagation model.

  • Step-3: the matrix is provided as input of the propagation model, so to find as output the estimated RSS measures in each area of the workspace. The model is resolved as follows:
    • Step-3.1: the matrix \(D\) is computed, where \(d(i,\,j)\) is the distance of the \(i\)-th point of the workspace from the \(j\)-th sensor on the direct path;

    • Step-3.2: the matrix \(A\) is evaluated, where \(a(i,\,j)\) is the attenuation of the signal due to the met obstacles, i.e., the second and third terms of the equation in Fig. 7;

    • Step-3.3: the matrix \(PL\) is estimated, where \(pl(i,\,j)\) is obtained using the equation in Fig. 7 considering the distance \(d(i,\,j)\) and the attenuation \(a(i,\,j)\);

    • Step-3.4: the matrix RSS is calculated, where \(d(i,\,j)\) is obtained by subtract to the output gain of the \(j\)-th sensor the value of \(pl(i,\,j).\)

  • Step-4: the RM is constructed from the RSS measures obtained from the model resolution.

Fig. 5

Process of obtaining an RM by using a model of the signal propagation

4.2 Optimal Placement of the Reference Sensors

In Fig. 4, the key element of Placement Path Finder is the Pattern Selection Solver, which searches within the set of all the possible deployment configurations looking for the optimal one. The issue is to find the suitable approach for such a duty.

As said before, the optimal configuration is the one that exhibits the best accuracy. It is quite obvious that, in general, increasing the number of sensors raises the accuracy. However, it also leads to an increase in the costs: having a lot of sensors causes higher static costs, and also implies a strong growth of dynamic costs. Specifically, we witness to a rise of the economic investment needed to acquire and deploy the positioning infrastructure (i.e., static costs), and the time needed to estimate the user location (i.e., dynamic costs). Static costs are paid only once when the system is built, so they are sustainable. On the contrary, dynamic costs strongly affect the efficiency of the positioning system since they are paid every time the system is used, and need to be properly managed. In particular, with respect to dynamic costs, there is an increase of either the time to collect RSS values from all the reference points (i.e., with respect to Eq. 2, having a large number of sensors increases \(T_{M}\)), either the time to query the RM in order to compute the user location (i.e., the fingerprint size grows when the number of sensors increases, so the size of RM and the time to perform a matching between measured fingerprint and stored fingerprint is higher). Even if the first increase is minor due to the low time to measure an RSS, the last one can not be neglected since the positioning system has to be loaded by mobile devices characterized by limited resources (both for processing and storing data). For this reason, the number of sensors can not exceed a certain threshold, otherwise the system cannot be properly used.

So, the number of sensors has to be chosen as small as possible in order to limit such drawbacks, but without compromising the accuracy provided by the system. Therefore, we can formulate the issue of deploying the reference points as a problem where accuracy has to be maximized while sensor number has to be minimized, under the condition of complete wireless coverage of the workspace (if there is no coverage in a particular area of the workspace, the system is not able to compute position in such an area). Such kind of problems are called Multi-Objective Optimization Problems (MOOPs), since they have the scope of simultaneously optimizing two or more conflicting objectives, subject to certain constraints.

Using a given solver to handle the issue of selecting the right number and location of the reference points, the steps performed by our approach to deploy RSS-based positioning systems are shown in Fig. 6:
  • Step-1: the user provides a graphical representation of the workspace, which is, then, converted in a matrix, as previously mentioned in Step-1 and Step-2 of the approach to build the RM. In addition, the user uses the Workspace Characterizer as follows:
    • indicating the wireless technologies that will be adopted within the workspace;

    • placing within the workspace representation sensors that cannot be moved at will (e.g., sensors already-deployed for communication concerns and used also for positioning purposes);

    • indicating the maximum number of sensors that can be deployed per each technology;

    • setting proper values for the inputs of the solver.

  • Step-2: the inputs of the user gathered in the previous step is provided to the solver, implemented in the Placement Pattern Finder, which returns a set of suboptimal solutions;

  • Step-3: the returned solutions are presented at the user, which decides the one to use. In fact, well-defined MOOPs do not have a single solution that simultaneously optimizes each objective, but a set of solutions for which each objective has been optimized to the extent that if it is further optimized, then the other objective(s) will suffer as a result.

Fig. 6

Methodology for deploying RSS-based positioning systems

5 Signal Propagation Model

As shown in Fig. 7b, a RSS pattern can be defined as the difference of the following two factors: Emitter Signal Strength (ESS), i.e., the output gain of the sensors used as reference points, which is provided by the transmitter manufacturer, and Path Loss (PL), i.e., the attenuation experienced by the signal on its path from the emitter to the receiver. As shown in Fig. 7a, a radio wave can take several routes, which generally fall into three distinct classes. Only one direct path exists and this is called Line-Of-Sight (LOS) if there are no obstacles, or, Obstructed Line-Of-Sight (OLOS) if one or more obstacles are placed between transmitter and receiver. In addition, as illustrated in Fig. 7, signals can reach a given destination by following other paths, caused by phenomena such as reflection. Such paths belong to the class called Non Line-Of-Sight (NLOS). To avoid becoming overwhelmed by attempting to consider all these paths, we have chosen to model signal propagation by means of the only OLOS. However, the methodology expressed by this paper can be applied even if the adopted signal propagation model is improved by considering multi-path effects. Despite exhibiting high accuracy, such models may not be suitable since they also imply high calculation complexity.
Fig. 7

Signal propagation: a Multiple paths that a signal can take to reach its destination, and b mathematical expression of RSS

We have adopted the so-called Multi-wall Model [33] to find the average signal strength at the receiver given some features of the transmitter, receiver and along the path, since it represents a good trade-off between accuracy and complexity. However, any other propagation model among the ones available in the literature [34, 35, 36] can be effectively used in our approach. Multi-Wall Model is based on the assumption that PL among obstacles is considered equal to the one in free space. So, PL is computed by considering the following contributions illustrated in Fig. 7b:
  1. 1.

    PL in free space depends on the logarithm of the distance, namely \(d\), between transmitter and receiver;

     
  2. 2.

    attenuation to pass through all the obstacles met along OLOS (where \(P_{i}\) is the number of obstacles of \(i\)-th type found on the OLOS, and \(AF_{i}\) is the attenuation factor associated to the \(i\)-th type of obstacle);

     
  3. 3.

    attenuation, namely \(FAF_{n}\), caused if the source and the destination are not placed on the same floor, which is function of the number of crossed floors, namely n;

     
  4. 4.

    attenuation, namely \(CoA_{i,j}\), produced by passive coexistence of two RF technologies, respectively indicated with indexes i and j, that use the same transmission frequency, namely \(\lambda\), and compete on the access to the channel, e.g., using WiFi and Bluetooth in the same workspace results in a potential for reciprocal interference [37].

     
The values for \(AF_{i}, FAF_{n}\) and \(CoA_{i,j}\) can be obtained by the literature, such as [38]; a concrete example is provided by Table 2. If they are not available in the literature or it is needed to have a better characterization of attenuation caused by elements within the workspace, they can be inferred from some on-field measurements. Specifically, such measurements can be performed without assuming a particular sensor deployment, but placing a certain AP behind the given obstacle and measuring RSS values at different locations, characterized by a certain distance and angle from the obstacles. Although such measurements are not complex to perform, they are very time-consuming. So, the evaluation of the model parameters by means of on-field measurements should be avoided as much as possible since they can limit the advantages brought by the application of our approach in terms of needed calibration time. Moreover, the detailed and exact modeling of attenuation factors of all the obstacles within the workspace is not crucial for a good RM and an optimal placement selection, as we have shown in Sect. 7.
Table 2

Typical attenuation factors for obstacles taken from [38]

Obstacle type

Loss (dB)

Moveable walls

1.4

Doors

2.0

Windows

2.0

Fixed walls

3.0

Metal partitions

5.0

Exterior walls

10.0

Basement walls

20.0

This model needs two corrections in order to achieve an acceptable accuracy. On one hand, the environmental pollution, multi-path or factors unconsidered into the model can cause an error between the simulated and the real PL. Therefore, we have changed the traditional Multi-wall model by adding an optional attenuation factor \(K\) in the equation (indicated as 5th addend in the formula shown in Fig. 7b), which represents the difference between the real and the simulated values. On the other hand, RFID sensors exhibit a peculiar characteristic that we do not meet when using WiFi and Bluetooth sensors: the adopted antenna is directional, i.e., the signal strength is higher in front of the antenna than behind or beside the antenna. Figure 8 provides a graphical description of this feature of directional antenna. Some studies, such as [39], have investigated the radiation area of this kind of antennas and concluded that it can be approximated by an ellipsoid, which is definable by the RFID system parameters. In other words, this approximation completely ignores the three tiny ellipsoids of Fig. 8, placed behind and beside the antenna. Despite of the lower accuracy achievable by using this approximation, we have decided to apply it in our approach due to its simplicity. However, this decision does not preclude an user to use any other RFID propagation model so to obtain an higher accuracy even at the cost of increasing model complexity. Therefore, in our work we assume that RSS for points outside the RFID radiation area is null, while inside it is computed by applying the previously-described Multi-wall Model.
Fig. 8

Signal strength of directional antennas, taken from http://www.connect802.com/wcu/2005/newsletter_050601.htm

6 MOOP Solver

In literature, there are several approaches to be used for solving MOOPs [40], and the most widely-adopted ones are the evolutionary algorithms [41] since they have been demonstrated to be general, robust and fast search mechanisms [42]. We have chosen a particular evolutionary algorithm known as Multi-Objective Genetic Algorithm (MOGA) [43], since it is simple to implement and efficient (at each iteration of the algorithm, more than one sub-optimal solution are generated and evaluated, so to reduce the convergence time), but any other evolutionary algorithm can be used.

MOGA, as any evolutionary optimization algorithm, works on two key elements: Chromosome, i.e., a representation of a certain solution to the given optimization problem, and Population, i.e., a collection of chromosomes that are analyzed at a given iteration of the algorithm. So, the description of the MOGA execution has to provide answers to two crucial questions: “How do solutions are represented by means of chromosomes?" and “How is the solution space explored by means of evolving from one population to another one?".

Figure 9a partially provides an answer to the first question. It illustrates how a chromosome is able to code a placement patter of sensors of the same RF technology. Specifically, a chromosome is made of a single vector of digits, named string, composed of two distinct parts: (i) a binary code, which represents the number of reference sensors, and (ii) a permutation vector, which contains the identifications of all the possible areas where a reference sensor may be placed. In particular, if in the placement pattern the number of reference sensors is set to \(n\), then the first \(n\) elements of the permutation vector represent where the reference points are located. Since we focused on deploying sensors of more than one RF technology, then a chromosome is composed by more than one string, each for a certain RF technology used within the positioning system. In addition, not all the chromosomes can be analyzed by the MOGA resolver, but only those, admissible, ones that verify the following rules:
  1. 1.

    The sum of the binary codes of all the strings does not have to exceed the maximum number of sensors indicated by the user;

     
  2. 2.
    It is impossible to find two permutation vectors that shares even a single number in its first positions. e.g., let us consider two permutation vectors, namely \(\bar{\alpha}\) and \(\bar{\beta}\), whose relative binary codes are respectively equal to \(n_{\alpha}\) and \(n_{\beta}\), the following condition is never satisfied:
    $$ \exists i \in \{0, \ldots , n_{\alpha}\} \wedge \exists j \in \{0, \ldots , n_{\beta}\} : \alpha_{i} \in \bar{\alpha}, \beta_{i} \in \bar{\beta} \Rightarrow \alpha_{i} = \beta_{j} $$
    (4)
     
Fig. 9

a Construction of a chromosome, and b execution of the MOGA resolver

An evolutionary algorithm, such as MOGA, explores the solution space of a given optimization problem as follows, depicted in Fig. 9b:
  • Step-1: chromosomes for the initial population are randomly calculated, based on user inputs;

  • Step-2: each chromosomes within the current population is evaluated by computing the relative RM through the approach presented in Sect. 5 and a quality score is assigned by estimating the accuracy of the positioning system with the computed RM and the relative signal coverage;

  • Step-3: the chromosomes that present greatest quality score and do not exhibit a positive value of coverage score is stored in an archive. If some of the chromosomes that are already stored in the archive exhibit worse quality scores than the new chromosomes, they are taken away from the archive.

  • Step-4: if the termination condition of the algorithm is verified, the content of the archive is returned to the user. Otherwise, the algorithm proceeds by executing the next step.

  • Step-5: at this point a new population needs to be generated, so to explore other possible solutions to the problem. Using the chromosomes in the current population, the new ones are determinated using the specific operators of biological evolution:
    • Mating new chromosomes are made by recombining the old ones. Two different techniques are applied between strings of the same RF technology for two mating chromosomes:
      • One-point cross-over for the binary part: A single cross-over point, namely \(\pi\), is selected. All binary digits beyond that point in either chromosomes are swapped between the two parent chromosomes. The resulting chromosomes are the children:
        $$ \begin{array}{ccc} \begin{array}{c} 010|1101 \\ 001|1001 \end{array} &\xrightarrow{\pi = 4} & \begin{array}{c} 010|1001 \\ 001|1101 \end{array} \end{array} $$
        (5)
      • OX cross-over for the permutation part: Two cross-over points, namely \(\pi_{1}\) and \(\pi_{2}\), are randomly selected. Everything between the two points is swapped between the parent chromosomes, rendering two child chromosomes:
        $$ \begin{array}{ccc} \begin{array}{c} 315|46|79 \\ 201|94|38 \end{array} & \xrightarrow{\pi_{1} = 4, \pi_{2} = 6} & \begin{array}{c} 201|46|38 \\ 315|94|79 \end{array} \end{array} $$
        (6)
  • Mutation a new chromosome is made by arbitrarily changing one value in the old one. Also in this case there are two different techniques depending if we have to mutate the binary or the permutation part of a string:
    • Bit-flip for the binary part: a point of mutation, namely \(\mu\), is randomly selected, and the binary digit in that place is changed to the opposite value, according to a given probability:
      $$ \begin{array}{ccc} 010|1101 &\xrightarrow{\mu = 4} &010|0101 \end{array} $$
      (7)
    • Reciprocal exchange for the permutation part: two points of mutation, namely \(\mu_{1}\) and \(\mu_{2}\), are randomly selected. The order of everything between the two points is inverted, according to a given probability:
      $$ \begin{array}{ccc} 315|46|79 &\xrightarrow{\mu_{1} = 4,\; \mu_{2} = 6} &315|64|79 \end{array} $$
      (8)

Given the new chromosomes, the algorithm executes the Step-2, by assessing their quality in terms of accuracy and coverage.

7 Experimental Evaluations

Based on the methodology introduced in the previous section, we have implemented a Java-based prototype5 that provides to a user the abilities of \((i)\) describing its workspace, \((ii)\) obtaining a RM for a given sensor placement pattern, and \((iii)\) computing a set of optimal placement patterns. We used this prototype to evaluate the quality of the proposed approach.

The experimental environment consisted of two pre-mises of the CINI laboratory of Naples: the old laboratory, which is composed of two areas located on two floors with a dimension of about \(270\,m^2\) and is shown in Fig. 10a, b, and the new laboratory, which has a dimension of about \(130\,\hbox{m}^2\) with a layout depicted in Figs. 11a and 12a. Several wireless devices electromagnetically pollute such a testing environment, and human presence has not been avoided. During our experiments, we have respectively considered the first laboratory decomposed in about 180 locations of about \(1.5\,hbox{m}\times1.5 \,\hbox{m}\), while the second one in about 60 areas of about \(1.5\,\hbox{m}\times1.5\, \hbox{m}\). Several measurement campaigns have been performed into these experimental environments in order to collect the real RSS fingerprints given a certain sensor placement. In particular, during each measurement campaign a different number of samples (they consist of a series of RSS measurements, one for each AP deployed in the workspace) were collected at each location. The average over the collected samples represents the RSS fingerprint at the given area of the workspace.
Fig. 10

Assessment of the proposed approach for RM construction: (a) location of the used WIFI APs displayed on the workspace map; (b) location of the used WIFI, Bluetooth and RFID APs displayed on the workspace map; (c) comparison of RM quality varying construction approaches and the number of measured RSS samples when only WiFi is used; (d) comparison of RM quality varying construction approaches and the number of measured RSS samples when WiFi, Bluetooth and RFID are used

Fig. 11

Assessment of the adopted signal propagation model: (a) AP location in the workspace map and direction along which the measurements have been performed; and (b) comparison of measured RSS values and estimated ones when varying the distance between measuring device and AP and the adopted RF technology

Fig. 12

Assessment of the proposed approach for placement selection: (a) used placement pattern displayed on the workspace map; and (b) evaluation of the placement selection outcomes

7.1 Signal Strength Model Evaluation

The scope of this subsection is to assess the goodness of the propagation model used within our calibration approach. For this aim, we have performed a measurement campaign by using one AP within the environment of the new CINI laboratory. We have placed it on one of the external walls and made measurements by augmenting the distance between the measuring device and the AP, by moving along the arrow depicted in Fig. 11a. We have repeated each measurement 20 times and made it varying the adopted RF technology. As illustrated in Fig. 11b, we registered an error between the estimated RSS values, namely \(RSS_{mod}\), and the measured ones, namely \(RSS_{meas}\), and it is expected. Specifically, the average error (i.e., the average of the difference between \(RSS_{mod}\) and \(RSS_{meas}\)) is respectively equal to 9,66 for WiFi, 0,47 for Bluetooth and 11,5 for RFID. The next two subsection will investigate how this imperfect modeling of the signal propagation influences RM construction and placement selection.

7.2 RM Construction Evaluation

In this subsection, we investigate the effects of the usage of a propagation model on the quality of an indoor RSS-based positioning system, and how it differs with respect to the case of building RMs by means of on-field measurement campaigns. As previously mentioned, more than one measure is taken per each area of the workspace, and more measurements are collected, less the average is affected by signal variability. So, it is needed a large number of samples per area to obtain RMs of high quality. But, the calibration time to construct an RM, i.e., \(T_{C}\) in Eq. 2 and consequently \(T_{D}\) in Eq. 3, is function of the time spent at each area to collect a certain number of measures and the total number of areas where measurements are performed, so a large number of samples that cover all the areas of the workspace implies an high calibration time. In our campaign, we have considered to reduce such measurement time, by making experiments where we have varied the number of collected samples at each area and the number of areas where measurements are performed. We have considered the solutions of performing measurements \((i)\) in a subset of the areas (whose elements are randomly chosen) and \((ii)\) the size of such subset is equal to the total number of areas divided by a factor, called sampling ratio and indicated as \(r\). We have assumed r respectively equal to \(1/2\) and \(1/4\). Moreover, we have also used an HMM-algorithm [8] to improve RM quality when measurements are not performed in all the areas (in this case the sampling ration is equal to \(1/2\)). We have compared such measurement-based approaches to our model-based one. As mentioned in Sect. 5, the signal propagation in a real indoor environment is subject to the diffraction, reflection, multi-path effects, noise and interference caused by the structure and the complexity of the environment, so we have applied to our model an attenuation factor K estimated of about 3dB. As depicted in Fig. 10, we have conducted two different campaigns in the old CINI laboratory: one using only WiFi sensors, and one with WiFi, Blueetooth and RFID sensors, whose deployments are illustrated in boxes A and B of the figure. Boxes C and D show the quality of the RMs achieved per each experiment. We define the quality of an RM as the ratio of cases with a right location identification, namely DetectedLoc and the number of locations, namely TotalLoc:
$$ Q = \frac{{DetectedLoc}}{{TotalLoc}} $$
(9)

As shown in figure, the quality provided by our approach is comparable with the best one that empirically builds a RM (i.e., HMM-based curve) considering about 35 training samples at each location and a sampling ratio equal to \(1/2\). Therefore, in order to get the same quality a measurement-based calibration would require about one hour and 45 minutes, while our method spend less than one minute. As shown in box D, when more than one technology is adopted in the workspace, the situation is exacerbated and more samples are required, increasing the time needed for a measurement-based calibration, while we have not witnessed a remarkable performance aggravation when using our model-based approach.

7.3 Pattern Selection Evaluation

As mentioned in Paragraph 2, there are no general guidelines or standard deployment patterns that we can use for evaluating our selection approach. We have decided to consider two placements of reference points (sensor patterns) belonging to three different technologies (i.e., Bluetooth (BT), WiFi and RFID), which are illustrated in Fig. 12a, and drawn from [44]: Saw-tooth Partitioning, which places the sensors according to a saw-tooth configuration, and Equal Partitioning, which decomposes the overall environment into four portions and puts a sensor of a technology at the 2.4 frequency (e.g., Bluetooth or WiFi) in the center of each partition and place an RFID sensor at the external walls. These two placements are the ones that someone would come up with if signal coverage and interference avoidance would be the only matters, without making use of any particular automated planning tool. Among the solutions generated by the tool, the placement pattern illustrated in Fig. 12a and indicated as “Approach Placement" has been chosen. Such a placement pattern has been compared with the previous two manual placement patterns. Specifically, we have deployed sensors according to one of the three patterns, and we have considered a fingerprinting RSS-based positioning system that estimates user location by matching the measured RSS and the one stored in the RM. Therefore, we have performed the experiments by randomly selecting several locations within the workspace (whose number is around 50), and we have used the positioning system around 20 times per each location so to obtain the user location in terms of identifier of one of the areas that partitions the workspace. We have defined as accuracy the distance between the real area where the user is located and the area estimated by the positioning system. Then, we have computed the average of such distances so to have a measure of the accuracy of the system. As depicted in the box b of Fig. 12, the placement pattern computed by our approach overwhelms the other two patterns since it presents better values for accuracy.

8 Conclusions and Future Work

Configuring hybrid RSS-based positioning systems is costly and time consuming, because an extensive measurement campaign is needed, and no general guidelines are available to choose which technology to use, how many sensors to deploy and where to locate them. Here, we have described a novel automatic tuning approach for positioning systems that provides twofold innovative contributions to address the issue of calibrating such positioning systems without requiring any heavy measurements campaign, but offering a rigorous method for the placement pattern selection. On one hand, we have proposed to build RM via a simulation model of the signal propagation, in order to compute the RSS patterns without an heavy on-field measurement campaign. Moreover, we have also described an approach to determine the better reference points configuration by formulating such issue as a multi-objective problem, and solving it by means of a multi-objective genetic algorithm. Experimental results presented in this paper proved that this approach is able to construct high-quality RM without requiring massive on-field measurements, and to select the optimal placement pattern in terms of lower accuracy offered by the positioning system deployed with this given placement.

Future work will be articulated in the following research activities:
  • In our work we have assumed that all the reference points used the same value for the transmission power. However, adjusting power allocation is an important solution to further optimize accuracy [45]. We have planned to extend our approach so to include the possibility of selecting the optimal power allocation.

  • We want to include more than one propagation model, so to investigate how a certain model affect the quality of the built RM, and the effectiveness of the selection process carried out by our tool. In addition, we want to study other MOGA solver, so to highlight the effects that the adopted solver has on the overall selection process.

  • The literature is full of more complex fingerprinting techniques rather than the simple matching between vectors of RSS values. We want to extend our approach for being able to completely describe such techniques and use them to emulate location estimation.

  • We intend to investigate how a model-based RM can be used to augment a measurement-based RM, so to reduce amount of required calibration data to a fraction without compromising accuracy.

Footnotes

References

  1. 1.
    E. Kaasinen, User needs for location-aware mobile services. Personal and Ubiquitous Computing, Vol. 7, No. 1, pp. 70–79, 2003.CrossRefGoogle Scholar
  2. 2.
    M. Hazas, J. Scott, and J. Krumm, Location-aware computing comes of age. IEEE Computer magazine, Vol. 37, No. 2, pp. 95–97, 2004.Google Scholar
  3. 3.
    Y. Gu, A. Lo, and I. Niemegeers, A survey of indoor positioning systems for wireless personal networks. IEEE Communications Surveys and Tutorials, Vol. 11, No. 1, pp. 13–32, 2009.Google Scholar
  4. 4.
    H. Liu, H. Darabi, P. Banerjee, and J. Liu, Survey of wireless indoor positioning techniques and systems. IEEE Transactions on Systems, Man and Cybernetics Part C: Applications and Reviews, Vol. 37, No. 6, pp. 1067–1080, 2007.CrossRefGoogle Scholar
  5. 5.
    J. Hightower and G. Borriello, Location systems for ubiquitous computing. IEEE Computer magazine, Vol. 34, No. 8, pp. 57–66, 2001.Google Scholar
  6. 6.
    E. Elnahrawy, L. Xiaoyan, and R. P. Martin, The limits of localization using signal strength: a comparative study. Proceedings of the First Annual IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks (SECON 2004), pp. 406–414, 2004.Google Scholar
  7. 7.
    C. L. Bowen and T. L. Martin, Combining position estimates to enhance user localization. Proceeding of the 9th International Symposium on Wireless Personal Multimedia Communications (WPMC06), pp. 648–652, 2006.Google Scholar
  8. 8.
    X. Chai and Q. Yang, Reducing the calibration effort for probabilistic indoor location estimation. IEEE Transactions on Mobile Computing, Vol. 6, No. 6, pp. 649–662, 2007.CrossRefGoogle Scholar
  9. 9.
    N. D. Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and A. T. Campbell, A survey of mobile phone sensing. IEEE Communications Magazine, Vol. 48, No. 9, pp. 140–150, 2010.CrossRefGoogle Scholar
  10. 10.
    C. L. Bowen and T. L. Martin, A survey of location privacy and an approach for solitary users. Proceedings of the 40th Annual Hawaii International Conference on System Sciences, p. 163c, 2007.Google Scholar
  11. 11.
    E. Aitenbichler and M. Mhlhuser, An ir local positioning system for smart items and devices. Proceedings of the 23rd IEEE International Conference on Distributed Computing Systems Workshops (IWSAWC03), pp. 334–339, 2003.Google Scholar
  12. 12.
    R. Want, A. Hopper, V. Falcao, and J. Gibbons, The active badge location system. ACM Transactions on Information Systems, Vol. 10, No. 1, pp. 91–102, 1992.CrossRefGoogle Scholar
  13. 13.
    S. J. Ingram, D. Harmer, and M. Quinlan, Ultrawideband indoor positioning systems and their use in emergencies. Proceedings of the IEEE Conference on Position Location and Navigation Symposium, pp. 706–715, 2004.Google Scholar
  14. 14.
    D. Focken and R. Stiefelhagen, Towards vision-based 3-d people tracking in a smart room. Proceedings of the 4th IEEE International Conference on Multimodal Interfaces (ICMI’02), pp. 400–405, 2002.Google Scholar
  15. 15.
    F. Raab, E. B. Blood, T. O. Steiner, and H. R, Jones, Magnetic position and orientation tracking system. IEEE Transactions Aerospace and Electronic Systems, Vol. AES-15, No. 5, pp. 709–718, 1979.CrossRefGoogle Scholar
  16. 16.
    S. Feldmann, K. Kyamakya, A. Zapater, and Z. Lue, An indoor bluetooth-based positioning system: concept, implementation and experimental evaluation. Proceedings of the International Conference on Wireless Networks (ICWN 03), pp. 109–113, 2003.Google Scholar
  17. 17.
    L. T. Son and P. Orten, Enhancing accuracy performance of bluetooth positioning. Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC 07), pp. 2726–2731, 2007.Google Scholar
  18. 18.
    S. C. Spinella, A. Iera, and A. Molinaro, On potentials and limitations of a hybrid wlan-rfid indoor positioning technique. International Journal of Navigation and Observation, special issue, Integrating Radio Positioning and Communications: New Synergies, 2010.Google Scholar
  19. 19.
    J. Bohn, ipos: A fault-tolerant and adaptive multi-sensor positioning architecture with qos guarantees. Sensor Review, Vol. 27, No. 3, pp. 239–249, 2007.CrossRefGoogle Scholar
  20. 20.
    S. Aparicio, J. Pérez, A. M. Bernardos, and J. R. Casar, A fusion method based on bluetooth and wlan technologies for indoor location. Proceedings of IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems, pp. 487–491, 2008.Google Scholar
  21. 21.
    J. Hightower, B. Brumitt, and G. Borriello, The location stack: A layered model for location in ubiquitous computing. Proceedings of the 4th IEEE Workshop on Mobile Computing Systems & Applications (WMCSA 2002), pp. 22–28, 2002.Google Scholar
  22. 22.
    A. Bekkali, H. Sanson, and M. Matsumoto, Rfid indoor positioning based on probabilistic rfid map and kalman filtering. Proceedings of the 3rd IEEE International Conference on Wireless and Mobile Computing, Networking and Communications (WIMOB 2004), pp. 21–26, 2007.Google Scholar
  23. 23.
    J. Krumm and J. C. Platt, Minimizing calibration efforts for an indoor 802.11 device location measurement system. Technical Report, Microsoft Research, 2003.Google Scholar
  24. 24.
    K. Sayrafian-Pour and D. Kaspar, A novel model-based indoor positioning using signal strength. Proceedings of the IEEE 18th International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC 07), pp. 1–5, 2007.Google Scholar
  25. 25.
    S. Y. Seidel and T. S. Rappaport, A ray tracing technique to predict path loss and delay spread inside buildings. Proceedings of the IEEE Global Telecommunications Conference (GLOBECOM 92), Vol. 2, pp. 649–653, 1992.Google Scholar
  26. 26.
    T. K. Sarkar, Z. Ji, K. K. A. Medour, and M. Salazar-Palma, A survey of various propagation models for mobile communication. IEEE Antennas and Propagation Magazine, Vol. 45, No. 3, pp. 51–82, 2003.Google Scholar
  27. 27.
    M. Vasquez and J.-K. Hao, A heuristic approach for antenna positioning in cellular networks. Journal of Heuristics, Vol. 7, No. 5, pp. 443–472, 2001.MATHCrossRefGoogle Scholar
  28. 28.
    K. Jaffrès-Runsera, J.-M. Gorceb, and S. Ubédab, Mono- and multi-objective formulations for the indoor wireless lan planning problem. Computers and Operations Research, Vol. 35, No. 12, pp. 3885–3901, 2008.CrossRefGoogle Scholar
  29. 29.
    M. Cardei and J. Wu, Energy-efficient coverage problems in wireless ad-hoc sensor networks. Computer Communications, Vol. 29, No. 4, pp. 413–420, 2006.CrossRefGoogle Scholar
  30. 30.
    M. Heidari and K. Pahfavan, Performance evaluation of indoor geolocation systems using propsim hardware and ray tracing software. Proceedings of the International Workshop on Wireless Ad-Hoc Networks, pp. 351–355, 2004.Google Scholar
  31. 31.
    R. Casas, D. Cuartielles, A. Marco, H. J. Gracia, and J. L. Falco, Hidden issues in deploying an indoor location system. IEEE Pervasive Computing, Vol. 6, No. 2, pp. 62–69, 2007.CrossRefGoogle Scholar
  32. 32.
    P. Kumar Ray and A. Mahajan, A genetic algorithm-based approach to calculate the optimal configuration of ultrasonic sensors in a 3d position estimation system. Robotics and Autonomous Systems, Vol. 41, No. 4, pp. 165–177, 2002.MATHCrossRefGoogle Scholar
  33. 33.
    D. Molkdar, Review on radio propagation into and within buildings. IEEE Proceedings of Microwaves, Antennas and Propagation, Vol. 138, No. 1, pp. 61–73, 1991.Google Scholar
  34. 34.
    H. Hashemi, The indoor radio propagation channel. Proceedings of the IEEE, Vol. 81, No. 7, pp. 943–968, 1993.Google Scholar
  35. 35.
    W. K. Tam and V. N. Tran, Propagation modelling for indoor wireless communication. Electronics & Communication Engineering Journal, Vol. 7, No. 5, pp. 221–228, 1995.CrossRefGoogle Scholar
  36. 36.
    P. Kumar Ray and A. Mahajan, Propagation prediction models for wireless communication systems. IEEE Transactions on Microwave Theory and Techniques, Vol. 50, No. 3, pp. 662–673, 2002.CrossRefGoogle Scholar
  37. 37.
    L. Ophir, Y. Bitran, and I. Sherman, Wi-fi (ieee 802.11) and bluetooth coexistence: issues and solutions. Proceedings of the 15th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC 2004), pp. 847–852, 2004.Google Scholar
  38. 38.
    R. Morrow, Wireless network coexistence. McGraw-Hill, New York, 2004.Google Scholar
  39. 39.
    R. Aliberti, E. Di Giampaolo, and G. Marrocco, A model to estimate the rfid read-region in real environments. Proceedings of the 38th European Microwave Conference, pp. 290–293, 2008.Google Scholar
  40. 40.
    R. T. Marler and J. S. Arora, Survey of multi-objective optimization methods for engineering. Structural and Multidisciplinary Optimization, Vol. 26, No. 6, pp. 369–395, 2004.MathSciNetCrossRefGoogle Scholar
  41. 41.
    M. Zitzler, M. Laumanns, and S. Bleuler, A tutorial on evolutionary multiobjective optimization. In X. Gandibleux et al., editors, Metaheuristics for Multiobjective Optimisation, Lecture Notes in Economics and Mathematical Systems, Vol. 535, pp. 3–37, 2004.Google Scholar
  42. 42.
    T. Bäck, U. Hammel, and H.-P. Schwefel, Evolutionary computation: Comments on the history and current state. IEEE Transactions on Evolutionary Computation, Vol. 1, No. 1, pp. 3–17, 1997.CrossRefGoogle Scholar
  43. 43.
    C. M. Fonseca and P. J. Fleming, Genetic algorithms for multi-objective optimization: Formulation, discussion and generalization. Proceedings of the 5th International Conference on Genetic Algorithms, pp. 416–423, 1999.Google Scholar
  44. 44.
    K. Maksuriwong, V. Varavithya, and N. Chaiyaratana, Wireless lan access point placement using a multi-objective genetic algorithm. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, Vol. 2, pp. 1944–1949, 2003.Google Scholar
  45. 45.
    K. Sayrafian-Pour and J. Perez, Robust indoor positioning based on received signal strength. Proceedings of the 2nd International Conference on Pervasive Computing and Applications (ICPCA 07), pp. 693–698, 2007.Google Scholar

Copyright information

© Springer Science+Business Media, LLC 2011

Authors and Affiliations

  1. 1.Dipartimento di Informatica e SistemisticaUniversitá di Napoli Federico IINapoliItaly
  2. 2.Laboratorio ITEM ‘C. Savy’, Consorzio Interuniversitario Nazionale per l’InformaticaNapoliItaly

Personalised recommendations