Localization, tracking, and imaging of targets in wireless sensor networks: An invited review

Authors


Abstract

[1] Wireless sensor networks (WSNs) have shown many attractive features in a lot of real-world applications that motivate their rapid and wide diffusion. One of the most challenging topics when dealing with WSNs is the localization and tracking of objects from measurements collected by the nodes themselves. Once distributed in a region without the knowledge of their positions, the nodes actively take part in the localization of the network as well as to the detection and monitoring of the presence and movements of targets lying within the sensed area. This paper reviews state-of-the-art systems and approaches developed for WSN-based localization and tracking of active as well as passive targets. The main focus is on systems that exploit the strength of the received signal, always available at the WSN nodes, without ad hoc or additional hardware. Recent strategies for WSN-based imaging are discussed as well.

1. Introduction

[2] Wireless sensor networks (WSNs) consist of networks of autonomous nodes, connected through wireless links, aimed at sensing the surrounding environment for monitoring physical parameters of interest [Akyildiz et al., 2002; Yick et al., 2008; Sohraby et al., 2007; Raghavendra et al., 2004]. Thanks to the reduced dimension of the nodes, their sensing capabilities, the low power consumption, the scalability of the network, and the affordable costs, WSNs have experimented a rapid and widespread diffusion. Originally developed for military purposes, WSNs are now effectively used in a wide range of real-world civil and industrial applications ranging from the environmental monitoring (e.g., precision agriculture [Beckwith et al., 2004; Prabhakar et al., 2007; Kwong et al., 2009; Martinelli et al., 2009; Hu et al., 2010] and surveillance of volcano eruptions [Werner-Allen et al., 2005], landslides, and avalanches [Ramesh, 2009; Kenney et al., 2009]) to home automation [Y. Li et al., 2009; Jin et al., 2008], from the monitoring of industrial processes [Madni, 2009; Korber et al., 2007] to the control of traffic or tunnels [Zou et al., 2009; Carmo et al., 2010] up to homeland security [Gaol et al., 2008; Jain and Vokkarane, 2008; Viani et al., 2009]. Since each node has usually short-range communication capabilities, WSNs are generally densely distributed in the environment and they can be profitably exploited for localization and tracking purposes. Indeed, besides sensing and monitoring issues, another key topic when dealing with WSNs is the definition of the positions and the tracking of static or moving objects either equipped (active/collaborative target) or not (passive target) with a node or a transponder [Boukerche et al., 2007; Patwari and Wilson, 2010; Li et al., 2002; Huang et al., 2009; Ou and Ssu, 2008; Ssu et al., 2005; Bachrach and Taylor, 2005; Shin et al., 2010]. In the “wearable” case, the object can be itself a node or with an onboard transponder wirelessly connected [Patwari et al., 2005; Latsoudas and Sidiropoulos, 2007].

[3] Among traditional localization techniques, the use of global positioning systems (GPSs) is not always the optimal solution because of the costs for having a GPS receiver at each node especially when multiple objects have to be tracked as well as for the limited spatial resolution. Moreover, GPS cannot be efficiently used for indoor applications and also its robustness against interferences is questionable. As regards collaborative (or “active”) systems, approaches analyzing some particular features of radio frequency (RF) signals (e.g., the time of arrival (TOA), the direction/angle of arrival (AOA), and the received signal strength (RSS)) have been successfully used to address the localization problem [Pahlavan and Li, 2002; Catovic and Sahinoglu, 2004]. The information from infrared (IR) [Hao et al., 2009; Zappi et al., 2010], acoustic [Chen et al., 2004; Merhi et al., 2009], or ultrawideband (UWB) [Oppermann et al., 2004; Zhang et al., 2009; Arik and Akan, 2010] signals has been also profitably processed. From an applicative prospective, a main advantage of radio waves lies in their ability to penetrate smoke, nonmetallic barriers and walls. In case of fire in a building, the dense smoke prevent the use of standard optical/infrared systems based on cameras.

[4] Although more complex than collaborative localization techniques, passive systems avoid ad hoc devices [Z. Li et al., 2009] that can cause some problems when dealing with people (e.g., privacy issues) and noncooperative subjects (e.g., old people or intruders) behind the increment of the costs. Indeed, wearable devices can undergo damages caused by voluntary or accidental reasons strongly affecting the reliability of the localization/tracking system. Such systems are based on the analysis of the variations of some physical quantities available at the WSN nodes and useful for solving the location and tracking problems. The measurement of the RF signals has been thoroughly exploited since the values of their descriptive parameters (e.g., the RSS) are available at the physical layer of each node without the need of any additional hardware. (The physical layer defines the means of transmitting the information/signals between the nodes of the network over the physical medium.) In this framework, several techniques have been proposed in state-of-the-art literature [Patwari and Wilson, 2010; Nakatsuka et al., 2008; Viani et al., 2008] to deal with many practical applications of interest (e.g., the possibility to locate people in buildings, determining their number, positions, and movements). As a matter of fact, the potential of radio waves to penetrate nonmetallic walls can be very useful also for building surveillance, monitoring, and tracking as shown by Wilson and Patwari [2010b]. Such an application can be extended to imaging problems [Pastorino, 2010]. In this case, the objective is not strictly related to the localization of one or more targets neither to determine its or their movements, but to image the whole investigation domain and its time/spatial variations. Although being a pioneering area of research, some studies have been already carried out. They are concerned with the definition of a “representation” of the attenuation of the RF links within a WSN infrastructured area [Wilson and Patwari, 2010a, 2010b] or the probability map of the presence of a target [Viani et al., 2010].

[5] Focusing on an environment infrastructured with a WSN, this paper is aimed at giving an overview of the state-of-the-art and current trends in localization, tracking, and imaging of targets. The main emphasis is given on RSS-based active and passive localization systems since they seem more suited for WSN applications because of the aforementioned reasons and as assessed by the scientific production on this subject as well.

[6] The outline of the paper is as follows. After a general mathematical formulation of the localization, tracking, and imaging problems in a WSN-infrastructured environment (section 2), the problems of localization and tracking of targets are formulated and addressed in section 3 where state-of-the-art techniques considering active (section 3.1) and passive (section 3.2) targets are summarized. Section 4 presents the approaches devoted to define an image of the investigation area extending the frontiers of the localization/tracking problem. Some conclusions on the role of WSNs as an enabling technology for localization, tracking, and imaging targets are drawn in section 5 where open problems are discussed and possible future developments envisaged.

2. Mathematical Formulation

[7] Let us consider a set of N nodes of a wireless sensor network distributed within a region D, called area under test or investigation domain, where one or more targets are still or moving. Let us refer to O as the total number of targets, whose number can be either known or unknown depending on the problem at hand. In the more general case, the targets can be objects, animals or people entering or moving in a reference scenario. Such targets (i.e., their physical properties) can be generally modeled by means of a spatial function χk (x, y), k = 1, …, O, (x, y) ∈ Dk, Dk being the support of the kth target where for simplicity, but without loss of generality, a 2-D scenario has been taken into account.

[8] As for the localization/tracking of the target, the problem at hand can be classified into two main categories: localization/tracking of cooperative targets and localization/tracking of noncooperative targets. In the former case (Figure 1a), the target is equipped with a node or a transponder (tag) that transmits/receives signals. The localization is called “collaborative” (or “active”) since the target itself takes actively part to the location estimation process [Patwari et al., 2005] and the target itself is a node of the wireless sensor network. In this case, the problem can be mathematically formulated as

equation image

where ζtot(xm, ym), m = 1, …, N − 1, are the field measurements at the N − 1 receiving nodes when the remaining one (i.e., the target) is transmitting, (xm, ym), m = 1, …, N − 1, being the position of the mth receiving node, Jkin (x, y) models an impressed source defined on the support of the target Dk, and equation image1 is the inhomogeneous Green function for the target-free configuration [Chew, 1990]. (An impressed source is supplied by an external power source (e.g., battery).) This implies that the data can be measured from the other N − 1 nodes only if the target is present in the environment [i.e., Jkin (x, y) ≠ 0, (x, y) ∈ D].

Figure 1.

Scenarios with (a) a cooperative/active and (b) a noncooperative/passive one where the emitted (continuous line) and scattered (dashed line) fields are represented.

[9] In the other case (Figure 1b), the targets are “device-free” and the detection problem is denoted as passive localization/tracking [Patwari and Wilson, 2010]. Accordingly, to sense such a scenario, the nth node generates a probing field ζninc (x, y) whose effects, due to the interactions with all the structures in D, are measured by the other N − 1 nodes. Mathematically, the relationship between measured data and target is

equation image

where χb (x, y), (x, y) ∈ Db = DDk models the known “background” described through an inhomogeneous function taking into account also the presence of obstacles, Jkin (x, y) is an equivalent source, and ζn,0tot is the field due to the field generated by the nth node without targets. (An equivalent source is induced on the target by virtue of the probing field ζninc (x, y).) Moreover, equation image0 is the free-space Green function [Chew, 1990].

[10] It is worth noting that the sum of the first two terms at the right-hand of (2) can be seen as the “incident” field illuminating the target. Therefore, the original target-detection problem is reformulated as the retrieval of the “differential” (with respect to the background) source Jkin (x, y), whose support defines the target domain Dk, starting from the knowledge of the field ζtot (xm, ym), m = 1, …, N − 1, measured at the N − 1 receiving nodes.

3. Localization and Tracking Through WSNs

[11] The aim of localization and tracking in WSN-infrastructured environment is the determination of the positions of the targets, equation imagek, k = 1, …, O, in D (localization problem, Figure 2a) and/or their trajectories (tracking problem, Figure 2b) from the knowledge of the field measurements (i.e., samples of physical quantities) available at a set of M reference nodes located at known and predetermined positions equation imagem, m = 1, …, M (MN) of the sensed area. In this case, neither the retrieval of Jkin (x, y), k = 1, …, O nor their supports/shapes Dk, k = 1, …, O are generally of interest. As a matter of fact, Dk is assumed to be a point coinciding with equation imagek. Therefore, simplified models taking into account the attenuation of the field strength due to the path loss or other propagation models have been proposed.

Figure 2.

(a) Localization problem and (b) tracking problem in an environment infrastructured with a wireless sensor network.

[12] Although localization and tracking are usually dealt with as distinct problems, the definition of a trajectory can be described as the solution of a set of localization problems at successive time instants where the position of the target is expressed as a function of time equation imagek (t), k = 1, …, O. The velocity of the target is then computed from the knowledge of the time interval Δt between two consecutive localizations of the target (i.e., equation imagek (t) and equation imagek (t + Δt)).

[13] From an algorithmic viewpoint, the main difference between the localization problem and the tracking one is that localization is an “one-time” detection procedure where the quality of the final solution (i.e., the accuracy of the estimation of the target location) is the only issue, while tracking is an “on-time” procedure where the fast processing is an additional constraint mandatory in real-time applications. The physical data to be collected at the WSN nodes for solving the two problems as well as the underlying modelings relating the solutions to the data are different. Consequently, not all WSN-based approaches designed for locating targets can be used for tracking purposes and vice versa. However, the different approaches mainly differ from the type of target (i.e., active/passive).

3.1. Cooperative Localization and Tracking

[14] In the framework of WSN-based cooperative localization and tracking [Li, 2007], different technological solutions have been proposed. The use of a dedicated hardware (e.g., radio frequency identification) [Tesoriero et al., 2009; Saab and Nakad, 2010] or the integration with already existing wireless local area network (WLAN) infrastructures [Rohrig and Kunemund, 2007; Lee et al., 2009] has been considered and different models as well as sensors have been proposed for determining the coordinates of targets within the domain under test. In the following, some representative solutions and architectures are described.

3.1.1. Infrared-Based Systems

[15] IR-based tracking through WSNs usually considers each target wearing a badge that periodically or on demand transmits a unique identifier in the infrared spectrum [Want et al., 1992]. A set of reference sensor nodes (called anchors) is positioned at fixed and a priori known locations to collect the badge identifiers. Such an information is then sent to a central server for establishing the locations of the targets within the investigation domain.

[16] Although providing an adequate estimation of the target position at low costs, the use of IR-based localization systems is until now and, to the best of the authors' knowledge, quite limited due to the short action range of the sensors, but especially, because of the difficulty/impossibility of infrared signals to penetrate smoke, obstacles, and obstructions.

3.1.2. Angle of Arrival–Based Systems

[17] Also to overcome the limitations of IR solutions, the use of alternative technologies has been explored. Among them, the estimation of the AOA of the signal emitted from an active target has been analyzed. In such a case, state-of-the-art processing techniques have been exploited [Van Veen and Buckley, 1988; Ottersten et al., 1993] since the receiving WSN nodes are generally equipped with arrays of at least two receiving elements as microphones for acoustic signals [Cevher et al., 2007] or antennas when processing RF signals. Other approaches adopt fixed or rotating directional antennas to determine the directions of arrival of the signals thanks to the generation of a narrow beam [Elnahrawy et al., 2007]. (A directional antenna is an antenna having the property of radiating or receiving electromagnetic waves more effectively in some directions than others.) The main drawback of AOA-based systems is the need for arrays of sensors or directional antennas [Balanis, 2008] with an unavoidable increase of the system complexity and of the costs of the WSN infrastructure.

3.1.3. Time of Arrival–Based Systems

[18] A different strategy aimed at determining the distance of the target and not the direction as for AOA system exploits the information coming from the time of arrival (TOA) [Lee and Scholtz, 2002] and/or the time difference of arrival (TDOA) [Merhi et al., 2009]. Both solutions can refer to RF signals or acoustics waves [Sivrikaya and Yener, 2004]. In the former case, the transmission instant and the propagation delay from the source to the receiving node are taken into account. However, the acoustic version is preferred to avoid complex TOA estimators and a precise synchronization among the sensor nodes as for RF signals by virtue of the limited speed of acoustic signals with respect to the electromagnetic ones.

[19] As far as TDOA localization systems are concerned, two different alternatives exist. The first studies the difference in time at which a single signal from the target arrives at three or more receiving nodes [Merhi et al., 2009]. The second one compares the delay between two signals of different nature (e.g., ultrasound and radio frequency) generated by the target at the same time instant. While ensuring good performances in outdoor environments without obstacles, the resolution accuracy highly reduces in indoor environments due to the multiple reflections of RF signals unless resorting to sophisticated signal processing techniques for mitigating the multipath fading [Kang et al., 2009].

[20] Thanks to the resolution capability due to the wide bandwidth as well as the possibility to penetrate building materials (e.g., concrete and wood), tracking systems based on ultrawideband signals have been successfully applied [Oppermann et al., 2004; Arik and Akan, 2010; Shi et al., 2005; Gezici et al., 2005]. The UWB transmitter benefits from a low complexity and reduced energy consumption. Moreover, the radiated signals with noise-like properties turn out to be robust to interferences and jammers. More specifically, an ad hoc system based on dedicated UWB sensors for indoor localization has been presented by Z. Li et al. [2009]. The system at hand is composed by three hierarchical levels: (1) a large number of simple tags for collecting energy from the surrounding environment and to transmit their identification codes, (2) a small number of hubs equipped with batteries and used as relay stations, and (3) few base stations directly connected to the power grid for calculating the tag locations and to control the network state. Since the tags do not have a receiver, they cannot work actively in the localization process. The tag localization is then based on the arrival time of UWB impulses transmitted from the tags to the reference nodes. The 3-D tag position is univocally determined whether at least four hubs receive the tag signal (while only three hubs are enough to identify the 2-D coordinates on a plane). The simultaneous and uncoordinated transmission from different tags can cause interferences and lead to reception errors especially in the presence of a high traffic intensity. In order to avoid such a drawback, a random strategy for determining the most suitable tag-transmission time has been proposed by Z. Li et al. [2009]. Moreover, it has been shown that a proper geometrical arrangement of the hubs can reduce the ambiguity in locating the targets [Z. Li et al., 2009; Bishop et al., 2010]. Such a feature can be further enhanced by formulating the nonlinear problem as the least square minimization of the following objective function

equation image

where equation imagek = (xk, yk, zk) and equation imagem = (xm, ym, zm) are the position of the target and the coordinates of mth hub where the TOA values are measured, respectively, tm is the clock time at the mth hub, tk is the tag transmission time, c is the speed of light, and wm is a real and positive weight related to the reliability of the TOA measure. Synchronization among the hubs is not required since at the beginning the positions and the transmission times of the four hubs are known. The position and the transmission time of each tag is then identified by exploiting the information coming from the hubs.

[21] Despite the effectiveness of the approach, the principal drawbacks are the need of a centralized system for processing the data from the hubs/anchors and their required high density distribution especially in noisy environments for a correct estimate of the target distance thus preventing their use in large sensor networks. To avoid these limitations, a distributed localization system has been proposed where the sensors and the anchors exchange data with their neighbors only [Khan et al., 2009a, 2009b, 2010; Zhang and Martonosi, 2008; M. Deghat et al., Distributed localization via barycentric coordinates: Finite-time convergence, manuscript in preparation, 2011]. It has been shown that the localization algorithm, based on the solution of a large system of linear algebraic equations where the system matrix is highly sparse [Khan et al., 2009b], converges to the exact target locations when noiseless data are available (Deghat et al., manuscript in preparation, 2011). Differently, when dealing with random environments (i.e., where the communication between two nodes can fail or noise is present), the approach leads to almost sure convergence in a limited number of iterations (Deghat et al., manuscript in preparation, 2011) through the minimization of a suitable cost function [Anderson et al., 2010; Bishop et al., 2009], also when the statistic and nature of the noise are completely unknown [Khan et al., 2010].

3.1.4. Received Signal Strength–Based Systems

[22] The advantage of RSS-based systems is mainly related to the use of data already available at the WSN nodes [Li, 2005; Klingbeil and Wark, 2008] without any additional hardware. Indeed, the RSS indicator is available at the physical layer of the node structure and it is not directly concerned with the “quality” of the signal, but to the received power [Saxena et al., 2008]. It is a number of 8 or 10 bits depending on the hardware of the WSN node and directly related to the accuracy of the tracking system. As for the spatial resolution, this is a drawback of RSS-based localization systems when compared to other techniques as, for example, acoustic TOA. Another disadvantage is the necessary knowledge of the radio propagation path loss model [Li et al., 2002] for enhancing the detection performances. As a matter of fact, this information is not available in many practical applications or it can be achieved only through an extensive and expensive set of experimental channel measurements. On the other hand, it is worthwhile to notice that alternative improvements have been proposed that avoid the channel modeling by resorting to algorithms adaptive to the changing environmental conditions [Li, 2006].

[23] Concerning RSS-based approaches that exploit a priori known (i.e., free-space conditions or determined through channel measurements) path loss models (5) to relate the received signal strength to the distance from the target, localization systems generally locate a target solving the following optimization problem (for simplicity let us refer to the two-dimensional (2-D) case)

equation image

where equation imagek = (xk, yk) is the unknown position of the kth target, γmk = equation image being the estimated distance of the kth target from the mth receiving node and dmk is proportional to the RSS value at equation imagem, RSSm, as follows [Pahlavan and Levesque, 1995; Cabrera-Mora and Xiao, 2008]

equation image

where P0 is the power loss (in dB) at the reference distance d0 (usually, d0 = 1m), Pt is the transmission signal power, α is the so-called path loss exponent, and ν is a Gaussian distributed random variable that models the shadowing effects of multipath environments with zero-mean value and variance σν. Starting from this model, Hara and Anzai [2008] presents the results of a comparative study in indoor environments on the performances of RF-based RSS and TDOA localization systems. The indoor scenario is not optimal for RF-based systems since RSS values could be unstable due to the multipath and the shadowing effect produced by the presence of both objects and targets. Therefore, one would expect that the use of alternative localization/tracking systems based on AOA, TOA, or TDOA should outperform RSS-based detections. On the contrary, it has been experimentally illustrated [Hara and Anzai, 2008] that the use of RSS signals is advantageous when the direct links, the so-called line on sight (LOS) of fundamental importance for TOA and TDOA systems, are often “obscured” by people/objects moving within the investigation domain.

[24] As for collaborative localization in WSNs, iterative approaches based on the maximum likelihood estimator (MLE) [Li, 2007; Koneru et al., 2006; Patwari et al., 2003] or on the multidimensional scaling (MDS) method [Latsoudas and Sidiropoulos, 2007; Li, 2007; Koneru et al., 2006; Costa et al., 2006; Shang et al., 2004] have shown satisfactory tracking capabilities also in complex scenarios.

[25] The MDS method performs the detection of the O targets, equipped with active devices as well as the M reference nodes, by minimizing the following function [Koneru et al., 2006]

equation image

where N = M + O and dij is the distance between the ith and the jth devices, γij is the distance between the same wireless nodes estimated through RSS-based path loss models, and wij is an integer coefficient equal to wij = 1 if γij is available or wij = 0, otherwise.

[26] As for the MLE, the cost function to be minimized for determining the targets locations is equal to

equation image

Although the formulation in (7) has been shown being more appropriate than MDS [Li, 2007] in fitting the statistical model of the data, the MDS solution turned out to be less sensitive to the initial estimate than the MLE solution [Koneru et al., 2006] with enhanced convergence properties. In order to exploit the positive features of both methods, a hybrid algorithm called MDS-MLE has been also formulated by Li [2007]. It consists of a two-step procedure where, at the first step, an initial estimate of the unknown positions is obtained by using the MDS to take advantage from its high convergence rate. The MLE is then applied at the second step to refine the solution and remove modeling errors inherent to the MDS method. The obtained results [Li, 2007] have shown the effectiveness of the hybrid approach as compared to each single-method solution. However, it is worth noting that the arrangement of the reference nodes plays a key role and it can positively/negatively affects the performance of the resulting detection accuracy whatever the algorithm at hand. As expected, the localization performances increase when the number of reference nodes grows [Latsoudas and Sidiropoulos, 2007; Li, 2007; Koneru et al., 2006]. Such an issue could represent a limiting factor for potential deployments in real-world applications.

[27] Unlike previous approaches and when it is not possible to estimate the path loss model through channel measurements [Li, 2006], as for monitoring and surveillance in hostile and/or inaccessible environments or when the channel characteristics drastically vary due to environmental or man-made causes, the target localization is carried out jointly with the estimation of the path loss exponent α on the basis of the RSS values [Li, 2005, 2006]. Although such a strategy appears more sensitive to the geometry of the infrastructure, the joint optimization is a viable solution in several realistic situations since classical techniques are very sensitive to the value of the path loss exponent that cannot be a priori carefully estimated.

[28] To improve the resolution accuracy as well as to mitigate the effects of the time-varying channel path loss, the use of simple preprocessing techniques has been investigated by Cabrera-Mora and Xiao [2008]. More specifically, it is assumed that the received signal strength is given by the superposition of two effects

equation image

sff (t) and ssf (t) being the fast and the slow fading, respectively. The slow fading is related to the attenuation due to the distance between transmitter and receiver. It takes into account electromagnetic phenomena such as reflection, diffraction, and scattering which are usually modeled with a lognormal distribution. The fast fading is concerned with the rapid fluctuations of the power signal due to physical factors such as the speed of the targets in time-varying scenarios or the bandwidth of the signal. When LOS conditions hold true, the fast fading is modeled with a Rice distribution, while a Rayleigh distribution is adopted when there is not a direct link between transmitter and receiver. However, it should be pointed out that the above distributions are not optimal for modeling the distribution of the signal values in many practical situations and especially for indoor environments. For such a reason, a method for the adaptive estimation of s(t) is presented by Cabrera-Mora and Xiao [2008]. It is based on the histogramic analysis. The target moving within D transmits, broadcast and at a fixed rate, its identifier. The receiver stores a set of RSS values for a fixed number of packets (e.g., 100 [Cabrera-Mora and Xiao, 2008]) and it determines a histogram of the RSS value occurrence. The RSS value with the maximum number of occurrences is chosen as reference to calibrate the propagation model between transmitter and receiver. Thanks to the histogram analysis, the large variations in the data associated to the fast fading are drastically reduced and the localization significantly improves even though the choice of the number of bins (i.e., the number of disjoint categories of an histogram) strongly influences the estimation accuracy [Song and Yu, 2008]. Setting this number is not trivial and it has to be carefully selected to ensure a suitable balance between memory storage and system performances [Cabrera-Mora and Xiao, 2008; Song and Yu, 2008]. To improve the robustness of the preprocessing technique, a compensation approach has been proposed by Cabrera-Mora and Xiao [2008] starting from the assumption that moderate changes of the distance between transmitter and receiver cannot cause a sudden variation of the RSS behavior. The approach is aimed at keeping constant the trend/behavior of the fading value with maximum occurrences by updating the current estimate when an abrupt change has been detected.

[29] A different solution to address the location and tracking problem when the received power is a complex function of the distance, as in indoor environments, is the use of learning-by-examples systems. Indeed, each indoor environment has some unique features in terms of signal propagation although the whole set belongs to the “indoor environment” class. Therefore, it is generally required to customize the detection system to the specific scenario at hand for obtaining reliable position estimates and no generalization capabilities are provided. Otherwise, learning-by-examples systems determine the signature of the investigation area during the initial process of training (performed offline) and then exploit such a knowledge to develop appropriate decision rules to be used during the (also real-time) testing phase (i.e., the localization process) also when dealing with similar (but not identical) scenarios. During the training, the features of the signal received at the network nodes are stored jointly with the (now) known position of the target to build a database of input-output relationships. At the test phase, pattern matching algorithms are applied to establish the unknown locations of the unknown target/s starting from the training database. Toward this purpose, both deterministic and probabilistic approaches have been used. Nearest neighbor classifiers and artificial neural networks (ANNs) [Ahmad et al., 2006; Stella et al., 2007] belong to the former class, while the algorithms based on the statistical learning theory to the latter one. In the work by Stella et al. [2007], the authors present a real-time ANN-based localization system which proved to be robust against noise and interferences with good generalization properties as confirmed by the satisfactory performances even in case of input data not belonging to the training set.

[30] A more classical approach based on k-nearest neighbors has been proposed by Ahmad et al. [2006]. The information on the received power is compared to the database values and the target position is determined by identifying the k values collected during the training (the nearest neighbors) with power much closer to the received one. The lack of scalability is unfortunately a severe limitation of such a technique.

[31] Still concerned with a data set of input-output examples, a regression-based approach has been presented by Yang and Chen [2009] together with an interesting correlation-based approach. While the first method is based on a linear regression between the RSS values and the target distance, the second one exploits the almost invariance of the RSS values from small displacements of the target position.

[32] For completeness, other approaches based on factor graph to subdivide the large scale global problems in a set of low-scale subproblems easier to solve [Huang et al., 2009] or Monte Carlo algorithm [Klingbeil and Wark, 2008] have been also analyzed.

[33] Besides methodologies aimed at determining the target position time-by-time, alternatives techniques propose the definition of a “target presence” area. This is the case of the method called Disk Overlapping [Liu et al., 2004] where the RSS values measured at different reference nodes are compared to determine the boundaries of the region where the target node should be with higher probability. Based on the hypothesis that if the RSS value at a sensor node location nA is greater than that in nB, then nA is with high probability closer to the target, the disk overlapping procedure works as follows. If the RSS value measured at the reference node nA is greater than that of node nB then the target is supposed to lie within the circle passing over nB and centered at nA [Liu et al., 2004]. The advantages are the range-free property and to avoid the estimation of path loss parameters during the detection process. On the contrary, the accuracy of the algorithm strongly depends on the number of reference nodes and their spatial distribution. In most cases, the mean square error of the estimate is higher than that of the MLE solution, but the performances of the disk overlapping can be improved through appropriate weighting strategies [Nabaee and Olfat, 2008] such that the points where the presence of the target is more likely have higher weight coefficients.

[34] As a final remark concerning all RSS-based systems, it is very useful to include the knowledge of the radiation pattern in the propagation model [Laitinen et al., 2007] since the variations of the receiver gain as a function of the angular coordinates can cause variations of the RSS value even though omnidirectional antennas are generally adopted since the orientation of the antenna with respect to the target is typically unknown.

3.2. Device-Free Passive Localization and Tracking

[35] Besides these approaches requiring the presence of a device on the target to transmit signals (of various nature) to the reference nodes of the WSN, another class of localization and tracking algorithms exists where the target is noncooperative and the target detection is performed by exploiting the signals generated by wireless sensor nodes or access points of a WLAN and the target itself [Pahlavan and Li, 2002]. Let us refer to these approaches as device-free localization (DFL) techniques. A key difference compared to collaborative localization systems is that the multipath effects do not negatively impact on DFL techniques, but, on the contrary, they can benefit from the presence of these phenomena [Patwari and Wilson, 2010]. Besides this feature, there has been a growing interest toward passive localization systems in recent years also because they do not require additional or ad hoc infrastructures and can be jointly used with other tracking systems (e.g., video cameras whose functionality can be limited by the presence of smog or the darkness of the scenario at hand, but in good conditions enable a high resolution accuracy) extending their sensing and monitoring capabilities. Most popular solutions process the amplitude-only RSS values available at the WSN nodes with low energy consumption and reduced costs. Otherwise, techniques based on UWB [Parekh and Cam, 2007; Paolini et al., 2008] and narrowband [Pahlavan and Li, 2002] signals have been also successfully used. The interested reader should also take into account that in addition to electromagnetic techniques (mainly considered in this overview), acoustic methods have been studied for estimating the location of a generic sound source as the position of a speaker in a room [Jia et al., 2009].

3.2.1. Infrared-Based Systems

[36] As for IR sensors, passive strategies have been proposed [Hao et al., 2009; Zappi et al., 2010; Song et al., 2008; Hao et al., 2006] where the localization procedure takes place when an object/person is detected in the “field of view” of the IR sensor. As an example, a hierarchical structure has been exploited by Zappi et al. [2010]. In such a case, several clusters of two nodes covering the investigation area are devoted to detect the direction of the movement and the relative position (i.e., close to one node, in the middle between two nodes, or close to the other node) of people. The “acquired” data are then transmitted to a central station for estimating positions and motions of the targets.

3.2.2. UWB-Based Systems

[37] UWB receivers measure amplitudes, temporal delays, and phases of the multipath components present in an indoor radio channel. The differences in the measurements taken at different time instants can be profitably exploited to determine both the properties of the static environment and the changes due to the movements of “targets”. Although UWB receivers are more expensive than the classical narrowband ones, UWB devices are able to sense the pulse response of the multipath channel modeled as [Hashemi, 1993]

equation image

where t and τ are the observation instant and the pulse-generation instant, respectively. Moreover, K(τ) is the number of multipath components, {αi (t)}, {τi (t)} and {θi (t)} are the amplitudes, arrival times, and phases of the received replicas, respectively. Under the assumption of a single bounce of the signal, the knowledge of the time delays among the multipath replicas gives sufficient information for determining the position of the target. This is possible without requiring the synchronization among the nodes although many approaches need the knowledge of the locations of the reference nodes (usually available) and the existence of the direct LOS component (in general, not always present in indoor environments). As a consequence, a main challenge of UWB localization lies into the definition of a robust postprocessing algorithm able to deal with the TOA of the LOS.

3.2.3. Narrowband-Based Systems

[38] Unlike UWB devices, narrowband receivers are quite cheap, but they only provide information on the amplitude and the phase of the voltage signal s(t) sum of all multipath contributions. More specifically, when a narrowband signal is transmitted in a highly multipath environment (e.g., an indoor installation), the measured phase does not provide an accurate estimate of the distance [Pahlavan and Li, 2002] since the received signal is the result of the interactions among various paths with different amplitudes and phases. Moreover, a time synchronization is necessary for reliable phase measurements, which is not easy in standard WSNs. Therefore, other WSN solutions based on amplitude-only measurements are usually preferred.

3.2.4. RSS-Based Systems

[39] Although the information coming from RSS values is less than the one coming from UWB signals, the reduced cost and the availability of the RSS at the network nodes without additional hardware make this solution very interesting by an applicative viewpoint. In the works by Youssef et al. [2007] and Moussa and Youssef [2009], the RSS measures collected by the nodes of a WiFi infrastructure have been used to detect the presence of targets and the extension of this approach to WSNs is of course straightforward. Since the spatial variability of the received RSS is typically due to multipath effects and to the changes of the distance between the transmitter and the receiver, while the temporal variations are mainly caused by the targets movements within the area of interest, the location strategy is based on the comparison between the long term behavior and the short term behavior. If the change is larger than a threshold, then an event is detected [Youssef et al., 2007]. Two different implementations of such a methodology have been discussed by Youssef et al. [2007] based on the computation of the average or the variance in a moving time window. In the first case, the average long-term and short-term coefficients at the jth time instant are computed as follows

equation image
equation image

βl and βs being the width of the long-term window and that of the short-term average, respectively. An “event” is detected if the condition

equation image

holds true, η being a user-defined threshold. Analogously, the deviation of the variance is considered to identify an event in the variance-based version of the comparative method [Youssef et al., 2007; Moussa and Youssef, 2009]. In both cases, the parameters setup is critical because of the dependence of the probability of false alarm on the user-defined thresholds. Moreover, the high variability of the power of the received signal in real indoor environments constitutes a very difficult benchmark especially for such a methodology.

[40] Still dealing with RSS-based systems, an approach for estimating the density of people has been analyzed by Nakatsuka et al. [2008]. The authors studied the relationship of the RSS value and of the link quality indicator, related to the number of correctly received information packets in a network, versus the number of people in the area under analysis. The guideline principle is the fact that the power associated to radio links is absorbed by people in the environment [Bultitude, 1987; Agrawal and Patwari, 2009]. It has been then observed that the mean value of the RSS decreases, while the variance grows, when the number of individuals between the two nodes increases [Nakatsuka et al., 2008].

[41] As for tracking, an effective RSS-based solution has been proposed by Wilson and Patwari [2010b] where an estimate of the target position is processed by a Kalman filter. The Kalman filter takes into account current and previous estimates to determine the target trajectory. Since objects are usually moving with a limited speed, a regularization is performed to keep down the noise effects and to prevent the tracking algorithm from the jump between nonadjacent positions.

4. Toward WSN-Based Imaging

[42] Imaging a scenario infrastructured with a WSN concerns with the solution of the inverse scattering problem which can be formulated through (2) where the target/s to be imagined is/are passive and embedded in an inhomogeneous background whose inhomogeneities are the other known objects besides the targets. Although the scenario can be probed by means of any kind of source, in the state-of-the-art literature either acoustic [Blomgren et al., 2002; Borcea et al., 2002] or radio [Wilson and Patwari, 2010a; Viani et al., 2010] signals have been mainly considered. Although the methods presented by Blomgren et al. [2002] and Borcea et al. [2002] have shown to effectively estimate the position of targets also when dealing with randomly inhomogeneous environments thanks to the superresolution [Blomgren et al., 2002] achievable by means of time-reversal refocusing strategies, customized approaches have been proposed [Wilson and Patwari, 2010a; Viani et al., 2010] for imaging environments through distributed sensing using WSNs.

[43] In the works by Wilson and Patwari [2010a] and Viani et al. [2010], the field used to probe the scenario under test is the electromagnetic field generated by the nth node [i.e., ζninc (x, y) = Eninc (x, y)]. In this case, the presence of obstacles and/or targets can be modeled in terms of so-called contrast function distribution

equation image

which is representative of the electrical properties of the media where χ (x, y) = ɛR (x, y) − 1 − jequation image, ω being the working angular frequency, ɛR and σ are the relative dielectric permittivity and the conductivity, respectively. Moreover, ɛ0 is the dielectric permittivity of the free space.

[44] From the equivalence theorem of the electromagnetic fields, it is well known that the field induced in D, ζtot (x, y) = Etot (x, y), is strongly related to that radiated in free-space by an equivalent current density [Ishimaru, 1991]

equation image

With reference to (2), the equivalent source is defined as Jkin (x, y) = χkin (x, y)Entot (x, y) where χkin (x, y) = χk (x, y) − χb (x, y) is the differential object function.

[45] It is worthwhile to note that solving such a problem allows one to determine the values of the complex source Jkin (x, y) not only its support, but also (in principle) the real and imaginary parts of the dielectric characteristics of the target domain. However, such an objective would require complex measurements (amplitude and phase) of the field, not available at the standard WSN nodes or in wireless access points. Moreover, computational expensive approaches would be necessary to define Jkin (x, y) through the minimization of a suitable cost function thus preventing real-time applications. To avoid this drawback different solutions have been proposed to obtain an image (Figure 3) of the area under test.

Figure 3.

Imaging problem in an environment infrastructured with a wireless sensor network.

[46] In the following, two different solutions will be discussed. The first one, named radio tomographic imaging (RTI) [Wilson and Patwari, 2010a], is aimed at defining a map of the attenuation of the links among the nodes of the network by means of the measure of the RSS values. The second technique is based on a learning-by-example method [Vapnik, 1998] and it looks for the probability that a portion of D is either occupied or not by one target [Viani et al., 2008, 2010].

4.1. Radio Tomographic Imaging

[47] The RTI approach is a DFL method devoted to detect the presence or/and activities and movements of targets (e.g., people) within an area under test [Wilson and Patwari, 2010a]. An RTI-based system exploits the idea that the links between pairs of nodes crossed by a target during its movement within a wireless network experience shadowing effects. An image of the attenuation of the wireless links can then be determined to infer the target location. More in detail, the investigation area D, where a set of wireless sensor nodes is present, is partitioned into C cells (Dc, c = 1, …, C), called voxels, such that D = equation imageDc. The RTI-based localization map is obtained by plotting the value of the power attenuation coefficient αc of each voxel due to the presence of one or more physical obstacles. The position of each target is then estimated from the location of the voxel/s where the power signal is more attenuated.

[48] As for propagation model, the path loss model used here is more complex than (5) since the RSS value associated to the ith link is given by

equation image

where Pt is the transmitted power, Li is the static loss due to the distance, Si (t) is the loss due to shadowing caused by targets, Fi (t) is the loss due to the multipath fading, and νi (t) models the measurement noise. The shadowing effect is modeled in terms of a weighted sum of the attenuation that occurs in a voxel crossed by the ith link as follows:

equation image

αc (t) being the attenuation at the cth voxel in the time instant t, while wic is a real coefficient. If the ith link does not cross the cth voxel, then wci = 0. Otherwise, an ellipsoidal weighing is applied by normalizing each coefficient wic to the length of the corresponding link.

[49] When time-varying phenomena are of interest, a RTI-based differential imaging strategy can also be applied [Wilson and Patwari, 2010a]. In such a case, the variations of the RSS are computed between two consecutive time instants (i.e., t and t + Δt),

equation image

where Γi = Fi (t + Δt) − Fi (t) + νi (t + Δt) − νi (t) and Δαc = αc (t + Δt) − αc (t). Since the set of linear equations (17) defines an ill-posed problem [Pastorino, 2010], a Tikhonov regularization has been proposed by Wilson and Patwari [2010a] to improve the location performances.

[50] Although quite promising, the resolution accuracy of the RTI technique is strongly influenced by the number of nodes. As a matter of fact, the granularity and the precision of the location system (i.e., the number of voxels) increases as equation image (N2) since each one of the N nodes can create a link with another node of the network. Despite the large number of nodes, RTI-based approaches enable the real-time tracking thanks to the reduced computational complexity. Moreover, they have shown to simply adapt to many different environments without expensive training phases.

[51] Other versions of the RTI technique analyze the variance on each link [Wilson and Patwari, 2010b] and, more recently, a new implementation that avoids the regularization step and is based on compressive sampling has been proposed starting from the hypothesis that few pixels of an area under test are usually occupied by targets [Kanso and Rabbat, 2009a, 2009b].

4.2. Support Vector Machine–Based Systems

[52] In the work by Viani et al. [2010], a learning-by-examples strategy based on a support vector machine (SVM) is assessed to address localization and tracking problems. The main motivation is the use of a flexible tool for real-time estimates also robust to the noise blurring RSS. Instead of directly determining the position, the size, and the shape of each target by imaging the corresponding equivalent current and successively determining the target support, the problem is formulated as the definition of the probability map that a target lies in a portion of D. The method has been implemented through a two-step procedure as follows. At the first step, a SVM-based classification is carried out. The region D is still divided into C cells and the cth cell has assigned a cell state, ϕc: ϕc = −1 when the cell is empty and ϕc = 1 when it is occupied by a target (i.e., the differential source Jin in the equivalent formulation). At the second step, the binary classification map is converted into a probability map by computing the probability that a cell is occupied, πc = Prc = 1∣(equation image, c)},

equation image

where equation image is the measurement vector containing the RSS data, p and q are curve-fitting parameters [Vapnik, 1998] and equation image (·) is a nonlinear function mapping the data space to a higher dimensional space (called feature space) where the two classes can be separated by a hyperplane equation image. It can be verified that the knowledge of equation image (·) is not required since one needs only to know its dot product in the feature space according to the so-called “kernel trick” [Massa et al., 2005]. The hyperplane equation image,

equation image

is determined (i.e., the unknown coefficients equation image and v denoting the normal vector and an offset) during the training phase that exploits the knowledge of a set of T known input-output relationships where the data and the corresponding classification maps are available.

5. Conclusions

[53] In this paper, a review of WSN-based techniques as applied to the localization, tracking, and imaging of targets has been reported. Representative state-of-the-art techniques for localization and tracking have been described by considering collaborative and device-free targets in an investigation area infrastructured with a WSN. Features, advantages, and drawbacks of each implementation have been discussed in a comparative fashion.

[54] The review has pointed out the following key issues.

[55] 1. Great care must be exercised in defining localization and tracking models because of the environmental conditions to be taken into account. With reference to indoor applications, the number of reflections caused by objects and walls unavoidably increases the complexity of the scenario at hand. The presence of multiple reflections has to be carefully considered and/or exploited when processing the measurements from the nodes or avoided through suitable strategies. As for outdoor applications, time-varying environmental conditions deeply influence the reliability of the systems and suitable countermeasures are needed.

[56] 2. The knowledge of the working conditions is fundamental to identify the most suitable technological solution to adopt for both signal generation and collection since optimal, general purpose techniques do not exist. That said, it should be noted that solutions based on RF signals have been widely used because of their capacity to penetrate many materials without being completely absorbed or reflected. In such a framework, systems based on RSS processing have had the widest diffusion.

[57] 3. The required resolution accuracy, the time constraints to get the estimation of the position/trajectory, the computational capability as well as the available data are all aspects that have to be taken into account to define/select a suitable localization algorithm.

[58] 4. The cost of a single sensor node, the probability of damages, the versatility of applications, the need of expensive calibrations, the solution scalability are issues to be carefully evaluated when designing and implementing a reliable localization/tracking system.

Acknowledgments

[59] The authors wish to express their gratitude to D. Lesselier.

Ancillary