Using mobile beacons to locate sensors in obstructed environments

J. Parallel Distrib. Comput. 70 (2010) 644–656

Contents lists available at ScienceDirect

J. Parallel Distrib. Comput. journal homepage: www.elsevier.com/locate/jpdc

Using mobile beacons to locate sensors in obstructed environments Yong Ding, Chen Wang, Li Xiao ∗ Department of Computer Science and Engineering, Michigan State University, United States

article

info

Article history: Received 11 February 2009 Received in revised form 26 February 2010 Accepted 4 March 2010 Available online 18 March 2010 Keywords: Sensor localization Ultrasound distance measurement Mobile beacon

abstract Locating sensors in an indoor environment is a challenging problem due to the insufficient distance measurements caused by short ultrasound range and the incorrect distance measurements caused by multipath effect of ultrasound. In this paper, we propose a virtual ruler approach, in which a vehicle equipped with multiple ultrasound beacons travels around the area to measure distances between pairwise sensors. Virtual Ruler can not only obtain sufficient distances between pairwise sensors, but can also eliminate incorrect distances in the distance measurement phase of sensor localization. We propose to measure the distance between pairwise sensors from multiple perspectives using the virtual ruler and filter incorrect values through a statistical approach. By assigning measured distances with confidence values, the localization algorithm can intelligently localize each sensor based on high confidence distances, which greatly improves localization accuracy. Our performance evaluation shows that the proposed approach can achieve better localization results than previous approaches in an indoor environment. © 2010 Elsevier Inc. All rights reserved.

1. Introduction Determining sensors’ locations is a critical task in the literature of sensor network research because it provides the fundamental support for many location-aware protocols and applications. For example, in location-aided routing protocols such as [11,31], the location information is critical for nodes to make optimal routing decisions. In sensor network applications, the sensed data labeled with the location is more informing because a geographic view of the data is available, such as [3,15,24]. As it is expensive to equip each sensor with a GPS, sensor localization usually involves two steps: (i) in-network distance measurement between pairwise nodes and (ii) geometric calculation based on measured distances. For the first step, radio signals and ultrasound are widely used as the distance measurement media. Because radio signals attenuate during their transmission, the transmission distance can be estimated from the received radio signal strength (RSS). However, the distance estimated by the RSS approach is often unreliable and inaccurate because radio signals are susceptible to environmental interference. The ultrasound approach estimates the distance between pairwise sensors by multiplying the time of flight (ToF) with sound’s constant speed. Accurate distance measurement can be achieved by the ultrasound approach if a line-of-sight path exists between pairwise sensors.

∗

Corresponding author. E-mail addresses: [email protected] (Y. Ding), [email protected] (C. Wang), [email protected] (L. Xiao). 0743-7315/$ – see front matter © 2010 Elsevier Inc. All rights reserved. doi:10.1016/j.jpdc.2010.03.002

Despite its accurate results in the line-of-sight condition, the ultrasound ToF approach must address two challenges before it can be readily applied to a sensor network deployed in a complicated environment, especially in an indoor environment where ultrasound signals are reflected along multiple paths. The first challenge is that the ultrasound ToF approach has short measurable range due to the power constraint of sensor nodes. The second challenge is that distance measurements will have large errors in an obstructed environment. As shown in Fig. 1, when the line-of-sight path is blocked between the sensors S1 and S2 , the distance has to be estimated from a reflected path which is much longer than the true distance between S1 and S2 . In this paper, we try to address these two challenges by using mobile beacons to measure distances between pairwise sensors deployed in an indoor environment. Fig. 2 illustrates the basic idea of our approach. We fixed a pair of ultrasound transmitters (or beacons) to the two ends of a vehicle. Assume that each sensor node is equipped with an ultrasound receiver in the network. In Fig. 2(a), the vehicle stops and lets the two beacons send out ultrasound signals. Thus, the two sensor nodes are able to measure their distances from both beacons, which are shown in solid lines. By knowing these four distances and the distance between the two beacons (the length of vehicle), we are able to calculate the distance between the pair of sensors, which is shown in the dashed line. As shown in Fig. 2(b), the vehicle moves through the deployed area, and stops at places for distance measurements. Finally, the sensors’ relative positions can be determined by all these measured distances. Note that this process is GPS-free, because we do not need to know the locations of beacons in order to calculate the distance between pairwise

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

645

quantified according to the distribution of its measured values. Based on the measurement confidence, the mobile distance measurement can be further combined with the recursive approach such that the distance measurement with higher confidence will have higher priority to be applied in the recursive approach. The rest of the paper is organized as follows. Section 2 summarizes previous work. Section 3 describes our experiments to evaluate RSS and Ultrasound ToF measurement in an indoor environment, especially the multipath effect of the ultrasound measurement. Section 4 describes the virtual ruler approach. Section 5 evaluates our proposed approach by comparing it with previous work. We conclude this paper in Section 6. Fig. 1. Ultrasound multipath effect in an indoor environment.

2. Related work

a

b

Fig. 2. The virtual ruler moves around the deployed area and measures pairwise distances between sensors in the network.

sensors. Here, the mobile beacons behave as a virtual ruler that wanders in the deployed area to provide distance measurement services to pairwise sensors. Compared with previous ultrasound based distance measurement approaches where a sensor acts as both a sender and a receiver, the virtual ruler approach can achieve longer distance measurement range such that more distance constraints are available to form a rigid network to uniquely determine sensors’ positions. Such a long range distance measurement can be achieved without violation of the energy constraints, because only beacons are equipped with high power ultrasound senders and sensors are equipped with receivers consuming less energy. As a result, the virtual ruler approach addresses the first challenge of short distance measurement range. To address the second challenge – distances estimated along the reflected paths have large errors – we conduct intensive experiments to test ultrasound distance measurement in an indoor environment. We observe that a sensor’s incorrect position, estimated from a reflected path instead of a straight line, is always mirrored to its true position. As shown in Fig. 1, the distance estimated between sensor S1 and S2 along the reflected path is equal to the distance between S1 and S20 , the mirrored position of S2 . Based on this phenomena, we further observe that incorrect distance measurements incurred by multipath effect have finite values that are virtual distances between sensors and their mirrors. This makes it possible to identify incorrect distance measurements through a statistical approach. When the virtual ruler moves around a pair of sensors, the distance between the pair of sensors can be measured by the virtual ruler multiple times from different perspectives. We observe two phenomena: (i) a distance between the same pairwise sensors can be measured by the virtual ruler more frequently if it is less affected by obstructions; (ii) among all the measured values to the same distance, the correct measurement value is observed more frequently than incorrect ones. Therefore, we can identify correct distance measurements based on the distribution of measured values. Moreover, the confidence to a distance measurement can be

Because knowing sensors’ positions is essential to many sensor applications, numerous approaches have been proposed to locate sensors; each approach has a different tradeoff between accuracy and cost. The localization approaches can be broadly categorized as radio based approaches and ultrasound based approaches according to their in-network distance measurement media. Due to its unreliable and inaccurate nature, the radio based RSS approach can only give a coarse estimation of distances between pairwise sensors [9,28]. To improve the accuracy of RSS method, the authors of [2] proposed a map-based localization algorithm, which learns a field map from sample radio signal strengths acquired in known positions during a preliminary environment exploration phase, and then uses the map in positioning. Although this approach improves localization accuracy, it needs a map learning phase, which is quite expensive, and the accuracy depends on the density of sampling. The radio based Time-of-Flight (ToF) method is more reliable than the RSS method, but requires stringent timing accuracy. In [13], the authors proposed methods for reducing error from clock offset and multi-path propagation. The approach is promising, but needs some additional hardware support. Instead of acquiring distance knowledge, radio signals are often used to estimate the connectivity between pairwise sensors, i.e. whether a pair of sensors are within the maximum radio transmission range. Based on the connectivity information obtained from radio signals, the Centroid approach has been proposed to estimate a sensor’s position as the centroid of the polygon formed by beacons that are connected to the estimated sensor through radio signals [4]. The accuracy of the Centroid approach has been further improved by the APIT approach that pinpoints the estimated sensor to a smaller area intersected by multiple triangles [7]. By approximating the hop count distance to the Euclidean distance between pairwise sensors, multihop approaches [18,21,22] were proposed to locate sensors in the case where beacons are sparsely distributed. Most indoor sensor applications require accurate localization results because sensors are deployed in a relatively small scale. In order for sensors to identify themselves based on their geographic coordinates, accurate localization results are necessary if they are densely deployed in a small area where distances between neighbors are short. Ultrasound based Time-of-Flight (ToF), which is more reliable and accurate than radio based approaches, is well suited to the indoor cases [6]. Methods of using ultrasound to measure distances between pairwise sensors have been extensively studied in [30,29]. By using ultrasound to measure distances from sensors to nearby static beacons, sensors’ positions can be determined with high accuracy. When beacons are sparsely distributed and not accessible to all sensors, the recursive approaches [20,1] have been proposed, which convert sensors to beacons after their positions

646

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

have been determined. Thus, beacons can be propagated in a deployed area and accessed by all sensors. To eliminate errors, especially the large flipping errors accumulated along the recursive approaches, the robust quad has been proposed to identify the situation where beacons are positioned close to a straight line [17]. All the research described above focuses on the localization algorithm using static in-network distance measurement and assumes distance measurements have small errors that can be tolerated by localization algorithms. In contrast, our work investigates how to use mobile beacons to dynamically measure distances in an indoor environment and proposes to eliminate incorrect distances estimated along reflected paths in the measurement step. Using mobile nodes in ad hoc wireless networks has been suggested by previous work [32,26]. Especially, using a mobile beacon to locate sensors has been proposed in [23], which assumes that the mobile beacon is equipped with GPS and knows its absolute coordinate. When the mobile beacon moves around a sensor, the sensor can estimate the distances to various positions of the mobile beacon. Based on its relative distances to the mobile beacon, the sensor’s absolute coordinate can be determined. Unlike this approach, our virtual ruler uses a pair of mobile beacons to measure distances between pairwise sensors instead of distances between sensors and beacons. The measured distances between pairwise sensors can be utilized by various localization algorithms to finally determine sensors’ relative coordinates. The virtual ruler approach does not require GPS to determine absolute coordinates of beacons, which is attractive for indoor localization where GPS signals are difficult to receive. The mobile-assisted approach proposed in [19] is similar to our work in that the beacons measure distances between pairwise nodes without reliance on GPS signals. Different from our work, the mobile-assisted approach uses a single mobile beacon to obtain distances between pairwise sensors. In order to have sufficient constraints to calculate the distance between pairwise sensors, the mobile-assisted approach requires the beacon to move along a certain track. For example, the mobile beacon must move along a straight line for two consecutive steps in the two dimensional localization. Moreover, the mobile-assisted approach pays less attention to the incorrect distance measurement incurred by obstructions. Based on distances acquired by mobile beacons, the iterative least squares fitting is proposed in [12] to exclude an incorrect distance measurement in each iteration when a measured distance significantly differs from the estimated distance. In contrast, our virtual ruler approach proposes to exclude the incorrect distances in the measurement step instead of the localization algorithm. Detailed performance comparison between the virtual ruler approach, the mobile assisted approach, and the iterative least square fitting approach is conducted in Section 4. Besides the previous ranging approaches, a radio interference based localization method has been proposed in [16]. It has been shown to be effective outdoors with large range of measurement and high accuracy. However, it currently cannot meet the localization requirements of indoor applications due to its sensitivity to multipath effect. The authors of [14] have shown that the indoor applicability of the radio interference based localization method still needs further theoretical and experimental research. In [8], the authors proposed to use the time of an event being sensed at each node to infer nodes’ locations. However, the system is designed for outdoors environment only. In a special application scenario, a large wireless sensor network is randomly deployed from a helicopter, and the sensor nodes localized themselves with the help of the helicopter to generate events over predefined traces. In this paper, we investigate applying the ultrasound ToF method in an obstructed indoor environment.

Fig. 3. Mica2 sensor node with wireless antenna.

3. Comparison of basic distance measurement methods in obstructed indoor environments In this section, we perform experiments to compare two commonly used distance measurement methods, RSS and ultrasound ToF, especially in an obstructed indoor environment. Compared with RSS, ultrasound ToF is a more accurate and reliable distance measurement method, because it suffers less from reflection and diffraction. However, we observe that ultrasound ToF method may still generate large errors due to the multi-path effect in the obstructed environment. This triggers us to use ultrasound ToF as the distance measurement media and employ techniques to eliminate these incorrect measurements. 3.1. RSS distance measurement In a Radio Signal Strength (RSS) ranging system, the sender transmits out radio signals, and the receiver measures the radio signal strength it receives. Ideally, the longer the distance, the weaker the signal strength that will be detected at the receiver. However, in an obstructed indoor environment, the reflection and diffraction effects of radio signals may significantly influence the signal strength detected at the receiver, which breaks the correlation between the distance between sensor nodes and the received signal strength. To study RSS distance measurement, we used MICA2 sensor nodes with wireless antennae (as shown in Fig. 3). The experiment results are illustrated in Figs. 5 and 6. The value of the signal strength is obtained from the RSSI pin on the radio signal module of the sensors. The RSSI value is inversely proportional to the signal strength. During distance measurement, the sender sends out messages continuously. The receiver collects the RSSI value of each message, and computes the mean for each 100 messages received. As shown in Fig. 5, in a near-ideal environment such as the open field, the received signal strength is roughly correlated with the distance between the sender and the receiver. As the distance increases, the received signal strength becomes weaker; that is, the RSSI value becomes greater. However, in our indoor environment setting where the receiver is placed near a wall, the received signal strength fluctuates in an unknown pattern with distance. This is because the reflected radio signal from the wall overlaps with the original signal, leading to strengthened or weakened signal at different points. In another experiment, we placed an obstacle between the sender and the receiver. The radio signal can still arrive at the receiver by diffraction or traveling through the obstacle. We measured the received signal strength when the sender and the receiver are placed at different distances apart with the obstacle in the middle. As illustrated in Fig. 6, the effect of obstructions can also disrupt the correlation between received signal strength and distance. We cannot uniquely judge the distance according to the received signal strength. We calculated the fitting function between the RSSI value and the real distance, and used the function to predict the distance

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

647

Fig. 4. Mica2 sensor node with ultrasound transceivers.

Fig. 6. Effect of obstruction on RSS distance measurement.

Fig. 5. Effect of reflection on RSS distance measurement.

from the measured RSSI value. We then compared the predicted distances with the real distances. For the case of measurement in open field, the RSS method achieves a relative error of 10.31% on average. For the case of receiver being placed at 8 ft from the wall in Fig. 5 and the case with an obstacle in Fig. 6, the RSS measurement causes a relative error of 17.9% and 15.6% on average, respectively. 3.2. Ultrasound distance measurement in line-of-sight condition Using ultrasound to measure distances in a sensor network has been extensively studied in previous work [6,29]. In this paper, we repeat the experiment of distance measurement in a line-of-sight condition and compare the result with that in an obstructed environment. We used the same hardware as the Cricket system [25], which attaches two ultrasound transceivers to the MICA2 nodes developed by UC Berkeley [10]. The Cricket sensor nodes (as shown in Fig. 4) broadcast the radio signals and ultrasonic signals at the same time. The ultrasonic signals will reach a receiving node later than the radio signals due to their speed difference. By measuring the flight delay of the ultrasonic signals, the distance between sender and receiver can be estimated based on the constant speed of ultrasonic signals. We measured the distances between a pair of Cricket sensor nodes in the line-of-sight condition. The pair of sensors were placed on platforms 2 in. above a hallway’s floor with ultrasound transceivers facing towards each other. We varied the distances between pairwise sensors from 0.3 ft to 30 ft. The experimental result is shown in Fig. 7, where the x-axis represents the actual distances and the y-axis represents the estimated distances. Each point shows an average of 20 measurements, and the variance for each point is below 0.0025 ft. Fig. 7 shows that a clear linear relationship exists between the actual distances and estimated distances. We can observe that the actual distances are slightly larger than the measured distances. This measurement error may be caused by several reasons. First, the speed of ultrasound depends on environmental factors such as temperature and humidity, and therefore is not constant. However, the speed of ultrasound is set as a constant value in Cricket nodes.

Fig. 7. Ultrasound distance measurement in line-of-sight.

Second, the clock of the sensors may not be accurate enough, which causes errors in the measurement of time. Third, sensor nodes may not be manufactured exactly the same as each other, and thus different sensors may use a slightly different time to process radio and ultrasound signals. We calibrated the pair of sensors through a fitting function between the actual distances and estimated distances (methods for calibrating sensors in large volume has been studied in [29]). The fitting function for Fig. 7 is y = 0.9 × x + 0.2. The linear coefficient corresponds to the first two types of errors, while the constant corresponds to the other type of error. After calibration, the ultrasound approach achieves a relative error of 0.22% on average, which is much more accurate than the RSS measurement (analyzed in Section 3.1). The reason that Ultrasound ToF is more accurate and reliable than RSS is that Ultrasound ToF measures the distance based on sound speed, which is less susceptible to the change of environment. The signal reflection and diffraction that influence RSS significantly have little impact on Ultrasound ToF method. 3.3. Ultrasound distance measurement in obstructed environments Despite its high measurement accuracy in the line-of-sight condition, the ultrasound approach may have large measurement errors in an indoor obstructed environment where multipath effects cannot be avoided. In an indoor environment, the ultrasonic signals may travel along multiple paths and arrive in a receiver at different time, though they start at the sender simultaneously. Nevertheless, if the line-of-sight path exists between pairwise sensors, the receiver can always identify the flight time along the shortest straight line by selecting the earliest arrival time of the ultrasonic signals. The situation becomes worse if a line-of-sight

648

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

Table 1 Comparison between the reflected distances and the distances from the sender to the receiver’s mirrored positions. Positions

Measured distance

Calibrated distance

Distance to mirrored position

Euclidean distance

m − n1 m − n2 m − n3 m − n4 m − n5

5.28 4.88 4.41 4.62 N/A

4.95 4.59 4.17 4.36 N/A

5.00 4.61 4.24 4.30 N/A

3.00 3.04 3.16 2.55 1.75

board

path does not exist between the sender and the receiver. As the straight line between a sender and a receiver is blocked by an obstruction, the ultrasonic signals have to travel a longer distance along reflected paths. Consequently, the estimated distance is much longer than the Euclidean distance between the pairwise sensors. We performed a series of experiments to study this phenomena. As shown in Fig. 8, we put the sender in the fixed position, while moving the receiver to different positions. We used a poster board as obstruction, which is place between the sender and the receiver. We verified that the ultrasound signal cannot travel through the obstruction. If the line-of-sight path between the sender and the receiver is blocked by the obstruction, the ultrasonic signals can travel along the path reflected by the wall. As the ultrasonic signals along the reflected paths have the same format as those along the straight lines, it is impossible for a single receiver to identify incorrect distance measurements affected by obstructions. However, our experiments show that a sensor’s false position estimated by the reflected path is always mirrored to the sensor’s true position. In other words, the distance estimated along the reflected path between pairwise sensors is exactly equal to the distance between the sender and the mirror of the receiver, which is shown in Fig. 9, where the true positions of the receivers are represented as circles and the mirrored positions are represented as crosses. In Fig. 9, the solid lines are the reflected paths from the sender m to the receiver’s positions n1 to n4 , and the dotted lines are the distances between the sender m to the mirrored positions of n1 to n4 . The receiver at position n5 cannot receive any ultrasound signals from m, because even the signals reflected by the wall are blocked by the obstruction. Table 1 lists the comparison results between the measured distances along the reflected paths and the calculated distances between the sender and mirrored positions of the receiver. Table 1 shows that the measured distances of the reflected paths are much larger than the Euclidean distances between the sender and the receiver but close to the distances between the sender and mirrored positions of the receiver. As the incorrect distance measurement between pairwise sensors is always equal to the virtual distance from one sensor to the mirrored position of the other, we can conclude the incorrect distance measurements incurred by obstructions have finite values since the mirrored positions of a sensor are finite given a finite number of obstructions. This observation motivates us to measure the distance between pairwise sensors from multiple perspectives and filter out incorrect distance measurements through a statistical approach. 4. Virtual ruler distance measurement approach As the studies in the previous section show that ultrasound ToF is a more accurate and reliable distance measurement method than RSS, we choose ultrasound as the distance measurement media in the obstructed indoor environment. However, several problems still exist with the ultrasound ToF method. First, it has a short range of measurement due to power constraints of

Wall receiver

sender Fig. 8. Experimental setup.

Fig. 9. Multi-path effect on ultrasound distance measurement.

n1 n2

d12 d11

d21 d22

m1

m2

(0, 0)

(0, L)

Fig. 10. Using virtual ruler to measure distance between sensors n1 and n2 .

sensor nodes (the maximum range is 10 m in our experiments). To increase the measurable range, higher transmission power is needed, which nevertheless contradicts the energy efficient design of sensors. Second, it may generate large errors when the line-of-sight is blocked by obstacles. To cope with these problems, we propose the virtual ruler approach in this section. Some preliminary results have already been published in [27]. In this paper, we further discuss the detailed localization algorithm design and evaluate it by more simulations. To overcome the short range problem, we attach two beacons on a vehicle, which travels around the whole network to measure pairwise distances between sensors. We also let the virtual ruler measure each distance from multiple perspectives to filter out incorrect distances statistically. In addition, our approach is GPS free. Below, we introduce a method that combines the virtual ruler approach with the recursive approach to localize all the sensors under the assumption of several known anchors in the network.

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

a

649

b

Fig. 11. Two possible distances.

4.1. Measure distances through virtual ruler In order to measure distances between pairwise sensors from multiple perspectives, we attach two beacons to a small vehicle that moves around in the deployed area. Because the distance between the attached beacons is fixed, we can easily infer the distance between a pair of sensors if the distances from the sensors to both beacons are known. Here, the mobile beacons behave as a virtual ruler that moves around to measure pairwise distances between sensors. Moreover, the distance between the same pair of sensors can be measured multiple times when mobile beacons move to different locations. Using mobile beacons to measure the distance between pairwise sensors is shown in Fig. 10, where nodes m1 and m2 are mobile beacons, and nodes n1 and n2 are sensors. As our purpose is to calculate the distance between n1 and n2, we can use relative coordinates for both beacons and sensors. Here, we assume m1 ’s coordinate is (0, 0) and m2 ’s coordinate is (0, L), where L is the fixed length of the vehicle. By measuring the distances dij from node ni (i = 1, 2) to beacon mj (j = 1, 2), the relative positions of sensors ni (i = 1, 2) can be calculated as below.

bi = arg min n ni

X (|ni − mj | − dij )2 .

(1)

Note that either n1 or n2 has two possible positions located at either side of the vehicle. Thus, there are four combinations of positions for both nodes, which result in two possible values of their Euclidean distance |n1 − n2 |. Fig. 11 illustrates the two possible distances obtained by using this method; we cannot decide which one is correct. However, this could be avoided by using the unidirectional transmission of ultrasound signals. We let the two beacons face to the same side of the vehicle, and thus the case of Fig. 11(b) can be eliminated. More robust results can be achieved when three beacons are laid out as a triangle on the vehicle. The measurement error of the virtual ruler is related with the length L between the two beacons. Higher distance measurement accuracy can be achieved by longer L. In our simulation, we set the L to 1 m, which makes it possible to attach the two beacons to a vehicle in real implementation. The ultrasound transmission range is 10 m and the ultrasound ToF method has a maximum error of 2 cm. We study the accuracy of virtual ruler in an unobstructed environment. We evaluate the accuracy by relative error, which is the ratio of the absolute error over the real distance. Fig. 12 illustrates the distance measurement error of the virtual ruler when the pairwise sensors are from different distances away from the beacons. The standard deviation of errors is also shown in the figure. Fig. 13 illustrates the maximum distance between beacons and sensors when the ranging system wants to achieve an accuracy within a certain error at the confidence of 0.9. For example, to ensure a relative error lower than 5% with a confidence of 0.9, the maximum distance between beacons and sensors can be 3.8 m. Although high accuracy can be achieved when line-of-sight paths exist from pairwise sensors to both beacons, the distance measurement may have large errors in an obstructed environment where some of the distances are measured along reflected paths.

Fig. 12. Accuracy of virtual ruler (cart length is set to 1 m).

However, as we discussed in Section 3.3, the incorrect distance measurements caused by obstructions have finite values. We illustrate this effect by simulations. We simulate two sensors in an obstructed environment, and let the virtual ruler measure their distance from different perspectives, where the distances from beacons to sensors are measured along reflected paths in places. The distribution of the measured values is shown in Fig. 14, where the y-axis represents the measured values and the x-axis represents the location from which the value is measured. From the figure, we can clearly see several clusters of values, among which the majority cluster corresponds to the real distance between the sensors. The distance measurement to the same pair of sensors has finite number of clusters because incorrect distances are equal to the virtual distances between a sensor to the finite mirrored positions of the other sensor. To further explain this phenomena, we list three examples in Fig. 15, where m1 and m2 represent the mobile beacons and n1 and n2 represent sensors, and the distance between n001 and n2 represents the incorrect distance estimation due to the multipath effects. Fig. 15(a), (b) and (c) show that 1, 2, and 4 line-of-sight paths are blocked by the obstruction respectively. An interesting observation is that when 4 line-of-sight paths are blocked by the obstruction in Fig. 15(c), both sensors are mirrored to the other side of the wall. As a result, the distance between the mirrored positions is equal to the distance between the original pairwise sensors. Therefore, the distance is correctly estimated by the mobile beacons, though they measure distances along the reflected paths to the sensors. When the virtual ruler moves around in the deployed area, it will measure distances between pairwise sensors after each moving step. During the movement, the virtual ruler can measure the distance between a pair of sensors from different perspectives to obtain different values, among which the correct distance measurement is mixed with the incorrect ones incurred by obstructions. In the following discussion, we show how to identify a correct distance measurement by assigning a confidence value to the measured distance. 4.2. Evaluate measured distances by virtual ruler In order to identify correct distance measurements, we conduct intensive simulations in various obstructed environments. When the virtual ruler moves around a pair of sensors, it can observe different distance measurement values d1 , d2 , . . . dn between the same pair of sensors from different perspectives. Moreover, the same value di can be measured by the virtual ruler multiple times from different locations. Let ki be the number of measurements of value di . Let N be the total number of measurements to the

650

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

Fig. 13. Accuracy of virtual ruler (with confidence of 0.9).

distance measurement, we define the former as the static beacons. Recursive approaches first locate sensors that are close to static beacons and iteratively convert sensors to static beacons after their positions are determined. Consequently, the static beacons can be propagated from the initial beacons to an entire deployed area such that all sensors can be localized. In the process of the recursive approaches, multiple candidates are often available to be converted to new static beacons. Distance measurements between pairwise sensors are also redundant in a densely deployed sensor network such that sensors can be localized by using only part of the distance measurements. The localization result will not be affected by the sequence of converting sensors to static beacons and the subset of measured distances if all the distance measurements have the same error distribution. However, in an obstructed indoor environment, where correct distances are mixed together with incorrect distances, it is critical to choose the optimal sequence of the recursive process such that the incorrect distance measurement can be excluded from the distance measurement subset used in the recursive process. Based on the confidence our virtual ruler assigns to distance measurement, the optimal recursive sequence can be approached as follows. In each step of the recursive approach, the candidate is selected such that we can maximize the confidence of all the distance measurements that are used to locate the candidate. 4.4. Localization algorithm design

Fig. 14. Distance measurement value distribution by virtual Ruler.

Pn

same pair of sensors. We have N = i=1 ki . Based on the distribution of the measured values, we observe two phenomena that are helpful in identifying correct distance measurements. First, a distance between a pair of sensors tends to have a larger N if it is less affected by obstructions. This is because fewer distance measurements are observed by beacons when the pair of sensors are surrounded by obstructions and therefore are difficult for the virtual ruler to access. Second, among all the distance values measured to the same pair of sensors, the value with the largest number of measurements kmax has the highest probability to be the correct distance measurement. Because a typical indoor environment contains more open spaces than obstructions, pairwise sensors have higher probability of being observed by mobile beacons through line-of-sight paths than through reflected paths. Based on the observations above, we choose the value with the largest number of measurements kmax as the correct distance measurement between a pair of sensors. Moreover, we assign confidence C to a distance measurement according to N and kmax as C = N + λ × kmax , where λ is the weighting coefficient. Based on the confidence of distance measurements, we combine the virtual ruler distance measurement with the recursive approach, in which the distance measurements with higher confidence will have higher priorities to be applied. 4.3. Combine virtual ruler distance measurement with recursive approach In the recursive approach, a few sensors are pointed as beacons whose positions are determined through out-of-band approaches such as manual measurements. In order to distinguish the beacons used by recursive approaches from the mobile beacons used by

When the virtual ruler travels around the area, it measures the distance between any two nodes multiple times at different places. We observe that if a pair of sensors are less affected by obstructions, their distance can be measured for a larger number of times. Therefore, in our localization algorithm, we prefer to use the distances with larger times of measurement. We may define a threshold n∗ such that we only use the distances that have been measured no less than n∗ times. On the other hand, we must have enough number of distances so that every node in the network has enough constraints to be localized. Algorithm 1 shows how this threshold can be determined, and thus we may prune the set of distances before localization. After we get the set of pruned distances, we may compute the position of each sensor by Algorithm 2. The basic idea is that each sensor tries to compute its position based on the distances with higher confidence. The algorithm goes through an iterative process. In each round, each sensor tries to discover distances with higher confidence to other nodes with known positions, and update its own position based on these distances. It terminates when no sensor can further update its position. In the algorithm, we use the multilateration approach to localize each node. As the focus of this paper is to eliminate the outlier in the distance measurement phase, we choose to use this simple and commonly used method in the experiment. The algorithm can be implemented in both centralized and distributed fashion. 1. Centralized implementation In the centralized implementation, the virtual ruler acts as a central point to collect all the distance measurements and compute sensors’ locations. When a virtual ruler is measuring the distance between a pair of sensors, its two beacons transmit out ultrasound signals and broadcast a message as well. The message contains only one field sid, which is used to uniquely identify this round of measurement. Each of the two sensors receives the ultrasound signal by its ultrasound receiver, and estimates its distances from the two beacons based on the ultrasound ToF approach. Thus, a record is generated for each sensor hsid, nodeid, d1, d2i, where nodeid is the unique identifier of the sensor, and d1, d2 are its distances from the two beacons respectively. Each sensor then sends this record to the virtual ruler. In this way, the

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

(a) Example a.

(b) Example b.

651

(c) Example c.

Fig. 15. Examples of distances estimated by virtual ruler in an obstructed environment.

Algorithm 1 Determine the Threshold of Distance Pruning 1:

2:

3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16:

Let A be the set of anchors, and N be the set of sensors that we need to localize. Let D be the set of measured distances. n(d)(d ∈ D) denotes the frequency of the distance being measured. Let L be the set of localized sensors and U be the set of unlocalized sensors. Let S be the set of selected distances, and T be the set of unselected distances. Initially, L = A, U = N, S = {}, T = D. while U 6= φ do Select d∗ from T such that n(d∗ ) = max{n(d) | d ∈ T } S = S + {d∗ } T = T − { d∗ } repeat for all n in N do if n can be localized (from S, n knows the distances to at least three nodes in L) then L = L + { n} U = U − { n} end if end for until no nodes in U can be localized end while n∗ = n(d∗ )

virtual ruler collects all the distance measurements after it has traversed the area. Then it is able to compute the distances between pairwise sensors measured from different perspectives, and compute sensors’ locations by Algorithm 2 finally. 2. Distributed implementation Algorithm 2 can also be implemented in a distributed way. Instead of sending the record to the virtual ruler during each measurement, each sensor keeps the record in its memory. After the virtual ruler has traversed the area, each sensor exchange the records with neighbors to calculate their pairwise distances. Then each sensor tries to localize itself based on the distances to its neighbors and their neighbors’ positions. In the distributed localization process, the anchors initially broadcast their positions to the neighbors. (i) If a sensor has not been localized yet, once it finds that it has more than three neighbors with known positions, it selects three of them to which it knows the distances with the highest confidence, and computes its own position. It then broadcasts its updated position to its neighbors. (ii) If a sensor has been localized, once it finds that any neighbor has updated its position, it selects three neighbors with distances of highest confidence. If the neighbors and their positions are different from those used in its previous position computation, it recalculates its own position, and broadcasts its updated position to its neighbors.

Algorithm 2 Localization Algorithm 1: 2: 3: 4: 5:

6: 7:

8: 9: 10: 11: 12: 13:

Let A be the set of anchors, and N be the set of sensors that we need to localize. Let D0 be the set of distances after pruning. repeat for all v in N do Find 3 distances with the highest confidence to 3 nodes in N ∪ A, whose positions are known (either anchors or sensor nodes that have been localized before). Denote the set of distances by D3 and set of nodes by V3 . if D3 and V3 could be found then if V3 is different from that used in previous localization of v OR the position of any node in V3 has been updated after the previous localization then Use D3 and the positions of V3 to recalculate the position of v . Update the position of v . end if end if end for until no node in N can update its localization result

4.5. Further discussions We assume that the virtual ruler moves around as below. It follows a step by step movement pattern. For each moving step, the virtual ruler randomly selects a moving direction. However, our test shows that such a random moving strategy leads to a nonuniform distribution of the moving tracks. As shown in Fig. 16, the virtual ruler only visits the right part of the floor while ignoring the left part. To improve the coverage of the virtual ruler, we assume that the virtual ruler has certain intelligence such that it can communicate with nearby sensors to decide the moving direction. During the process of measurement, each sensor keeps recording how many times it has been measured. The virtual ruler queries neighboring sensors before it makes the moving decision. Based on the queried results, the virtual ruler moves towards the sensor that has the least measurement times. This moving strategy requires some additional hardware support. The virtual ruler should be capable of measuring the direction of any sensor from which it receives the radio signal (angel of arrival), and controlling the direction of movement. The moving track of the enhanced moving strategy is shown in Fig. 17, where the floor is uniformly covered by the virtual ruler and all the sensors have been visited. Another problem that we should deal with in a real system is the uni-directional nature of ultrasound. According to our experiment with the Cricket sensor node, its ultrasound transmitter already scatters ultrasound signal to different directions, which forms a cone with degree over 60. In a real system, when the vehicle stops for measurement, we can tune the ultrasound transmitter of the

652

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

Fig. 16. Virtual ruler’s moving tracks under random moving strategy.

Fig. 17. Virtual ruler’s moving tracks under enhanced moving strategy.

Fig. 18. Sensors deployment with 20 randomly distributed obstructions.

beacon towards different directions for measurement. Moreover, it would be more convenient if the ultrasound could be scattered to a larger range by adding some additional hardware support in the transmitter. 5. Performance evaluation In this section, we evaluate our virtual ruler approach by simulation in MATLAB. We assume that the ultrasound signal has a maximum of 10 m range, and the distance measurement of ultrasound ToF method has a maximum error of 2 cm. The ultrasound signal can be reflected by obstacles. In our simulation, we set a 20 m × 20 m square area where 50 sensors are randomly deployed. Two configurations of the deployed area are used: (i) 20 obstructions are randomly positioned (Fig. 18); (ii) the square area is configured into a real indoor environment where rooms are separated from each other by walls (Fig. 17).

The virtual ruler moves around in the area and measures pairwise distances of static sensors within their measurable range. As illustrated in Figs. 17 and 18, the dashed lines show the trace of the virtual ruler that moves throughout the region under the enhanced moving strategy. The virtual ruler moves a total of 100 steps with step length of 4 m, and stops at each step to perform measurements. We can put the virtual ruler on a robot with an average speed of 0.8 m/s [5]. At each step, the virtual ruler stops for 1 s for distance measurements. Thus, the total time taken for measurement is around 10 min. The localization algorithm runs less than 2 min on a laptop computer with 2GH CPU. Therefore, the total time taken for localization is less than 12 min. This is acceptable for static wireless sensor networks, because the localization process is usually performed only once after the initial deployment of the sensor network. After deployment, the sensors rarely change locations. In this section, we evaluate both the distance measurement performance and the localization performance when combining the virtual ruler approach with the recursive approach. 5.1. Distance measurement performance of the mobile beacon approach When the virtual ruler moves around in the deployed area, it can measure the distance between a pair of sensors multiple times from different perspectives. For the distance between the same pair of sensors, the virtual ruler may obtain different values, among which the correct value is mixed together with the incorrect ones. In order to statistically identify the correct value from the incorrect ones, we investigate the characteristics of the distance measurement distribution through intensive simulations conducted in indoor environments. We record all the distance measurement values obtained by the virtual ruler and plot the distribution of measured values in Fig. 19, where Fig. 19(a) shows the full distribution of all the distance measurements and Fig. 19(b) shows an enlarged part of Fig. 19(a) for clear visualization. The height of each vertical bar represents the total number of distance measurements N between a pair of sensors. The vertical bar is further divided into two segments and the height of the the bottom segment represents the number of measurements kmax of the value dmax , which has the largest number of measurements among all the values observed by the virtual ruler. The sum of the number of measurements of all other values is represented as the length of the top bar. The bottom bar is colored in gray if it is the correct value; otherwise the bar is colored in black. Fig. 19 shows that the majority of bottom bars are painted in gray, which means the correct distance measurements tend to be observed by the virtual ruler more frequently than incorrect ones. Therefore, among all the measured distance values to the same pair of sensors, we can select the one with the highest observed frequency as the correct distance measurement. We compare the distance measurement of the mobile-assisted approach [19], virtual ruler approach with one time measurement, and virtual ruler approach with frequency analysis in both the indoor environment and random obstruction environment. The comparison results are shown in Tables 2 and 3 respectively. The comparison shows that the virtual ruler approach with one time measurement has lower percentage of incorrect distance measurement than the mobile-assisted approach. Moreover, the percentage of incorrect distance measurements of the virtual ruler approach can be significantly reduced by frequency analysis. From Fig. 19, we can find a few counter examples where incorrect distance measurements fall to the bottom bars that are painted in black. However, the bar that contains the correct distance measurement at the bottom tends to be high. This observation motivates us to further exclude incorrect distance measurement by imposing a threshold to distance measurement. When the total

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

(a) Full distribution.

653

(b) Enlarged part. Fig. 19. Distance measurement distribution.

Table 2 Distance measurement comparison in an indoor floor environment. The distance measurement approach

Mobile-assisted

Virtual ruler with one time measurement

Virtual ruler with frequency analysis

Virtual ruler with threshold strategy

Total number of measurements Number of incorrect ones Percentage of incorrect ones

486 129 26.54

658 146 22.19

658 100 15.2

201 2 1

Table 3 Distance measurement comparison in an area with randomly distributed obstacles. The distance measurement approach

Mobile-assisted

Virtual ruler with one time measurement

Virtual ruler with frequency analysis

Virtual ruler with threshold strategy

Total number of measurements Number of incorrect ones Percentage of incorrect ones

579 159 27.46

740 166 22.43

740 104 14.05

253 3 1.2

number of distance measurements falls below the threshold, we do not use the distance in the localization. By using Algorithm 1, we can determine the threshold to eliminate some distances with lower number of measurements while keeping enough distances for localization. As shown in Tables 2 and 3, the percentage of incorrect distance measurement is further reduced by using the threshold strategy. 5.2. Localization performance by combining the mobile beacons distance measurement with recursive approaches The observations above show that a distance measurement can be evaluated through two metrics: (i) the total number of measurements to the same distance between pairwise sensors; (ii) among all the measurements, the proportion of the value that has the largest contribution to the total number of measurements. Based on the two metrics, a confidence value can be assigned to each distance measurement such that we can rank distance measurements according to their confidences. With the ranked distance measurements, the virtual ruler distance measurement can be readily combined with the recursive approach to exclude incorrect distance measurements by giving a higher priority to a distance measurement with higher confidence. We then evaluate the combination of the virtual ruler approach and the recursive approach in both the environment with randomly distributed obstacles and the real indoor floor environment. We first used the centralized implementation of Algorithm 2 in our simulation. The simulation results for the random obstacles case and real environment case are shown in Figs. 20 and 21. Fig. 20 shows that our approach successfully selects a subset of the distance measurements with high confidence, which is enough to localize each sensor node in the network. In the random obstacles case, only three incorrect distances (painted in bold dashed lines) are used in location computation, while in the real environment case, only two incorrect distances are used. Fig. 21 illustrates the

Table 4 Localization error in an indoor floor environment.

Random obstacle Real floor plan

Minimum error (m)

Maximum error (m)

Median error (m)

Mean error (m)

0.29 0.21

0.88 0.72

0.42 0.40

0.50 0.43

localization error of each sensor node, where circles represent the true positions of sensors and lines represent the localization errors between true positions and the estimated positions. We repeated the simulation for 20 times. The minimum, maximum, median and mean error for each environment are shown in Table 4. Our simulation result shows that the average localization error of sensors is 0.50 m and 0.43 m respectively. We also simulated the distributed version of Algorithm 2, and it is able to get the same localization accuracy with the centralized version. Fig. 22 illustrates the average number of location recalculations on each sensor under different network scales. We can observe that the distributed algorithm converges quite fast, and does not require much computation overhead on each sensor. The only type of message exchanged is the local broadcast message of a sensor’s position. The number of messages broadcasted by each node is equal to the number of its location recalculations, because a sensor broadcasts its position only when it has been updated. Therefore, the overhead of the distributed algorithm is acceptable in a sensor network. 5.3. Comparison with other approaches In this section, we compare the virtual ruler approach with the iterative least squares fitting algorithm in [12] that filters out incorrect distance measurements by iteratively applying least squares fitting. In each iteration the estimated distances between a sensor and static beacons are calculated based on the sensor’s position estimated by least squares fitting. The

654

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

(a) Random obstacles.

(b) Real environment. Fig. 20. The subset of distance measurements selected by virtual ruler.

(a) Random obstacles.

(b) Real environment.

Fig. 21. Localization error by combining virtual ruler with recursive approach.

distance measurement that differs most from its estimated value is excluded in the next iteration. In addition, we also compare our approach with a variation of the above algorithm by using studentized residual. Fig. 23 compares the results of the different approaches in the real environment. The horizontal axis illustrates the Gaussian random error in the basic ultrasound ToF distance measurement, and the vertical axis shows the average localization error. We can see that our approach outperforms the iterative least squares fitting algorithm. The iterative least squares fitting algorithm fits all the measured distances equally such that the final estimated location is the averaged result of all the measurement. However, the least squares fitting algorithm does not guarantee that the estimated result will favor the correct distance measurements. If the estimated location favors the incorrect distance measurements, the correct distance measurement will be filtered in the iteration. In Fig. 23, we can observe a converging trend of the iterative least square fitting and our approach. This can be explained as below. Basically, there are two types of errors in distance measurement that affect the performance of sensor localization, that is, the Gaussian distributed random error in ultrasound measurement and the incorrect distance measurement caused by multipath effect. The second type usually possesses large errors, which is more harmful to sensor localization. Compared with Iterative Least Square Fitting, our approach is more efficient in eliminating the second type of error. However, when the first type of error increases and becomes dominant in the total error, the performance of the two approaches will converge. In real applications, the Gaussian random error in ultrasound measurement is quite small. Therefore, our approach is able to achieve better performance in indoor sensor localization.

Fig. 22. Computation overhead of distributed localization algorithm.

5.4. Impact of obstacle density and sensor density We perform extensive experiments to study the effect of the density of obstacles on the localization error. We vary the number of random obstacles in the area, and use the virtual ruler approach to measure distances and select high confidence distances for recursive localization. The result is shown in Fig. 24. We can observe that when the density of the obstacle is high, the chance of a correct distance measurement becomes very small, and thus the performance of our approach degrades accordingly. However, this is a very hard problem in the state-of-art research, because it becomes impossible to eliminate most of the incorrect distances under such extreme conditions. In this paper, we have focused on

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

655

mixed together with correct distance measurements, are difficult for localization algorithms to identify and exclude. In this paper, we proposed to filter out the incorrect distance measurement in the first step of distance measurement. By using mobile beacons to measure distances between pairwise sensors from multiple perspectives, our proposed mobile beacon based distance measurement can statistically identify incorrect distance measurements, which provides a good basis for indoor localization algorithms. Especially, the mobile beacon based distance measurement can be further combined with the recursive approach such that distance measurements with higher confidence are selected with higher priority. The performance evaluation shows that the mobile beacon based distance measurement can achieve better localization results than previous mobile-assisted approaches. Fig. 23. Compare virtual ruler with iterative least squares fitting.

Acknowledgments The authors would like to thank the anonymous reviewers for their helpful comments. This work was supported in part by the US National Science Foundation under Grants CCF-0514078, CNS0551464, and CNS-0721441. References

Fig. 24. The effect of obstacle density on localization performance.

Fig. 25. The effect of sensor density on localization performance.

2D simulations. It would be a prospective future research area to analyze the virtual ruler performance in large-scale test-beds with 3D shape obstructions. We also repeat the simulation with different number of sensors in the deployed area with 20 random obstacles. The localization accuracy of our virtual ruler approach is shown in Fig. 25. The results show that our approach is applicable to variable sensor densities. 6. Conclusion It is a challenging task to locate sensors in an indoor environment because the multi-path effects will incur large errors in distance measurements. The incorrect distance measurements, once

[1] J. Albowicz, A. Chen, L. Zhang, Recursive position estimation in sensor networks, in: ICNP, 2001. [2] C. Alippi, A. Mottarella, G. Vanini, A rf map-based localization algorithm for indoor environments, in: ISCAS, 2005. [3] J. Aslam, Z. Butler, F. Constantin, V. Crespi, G. Cybenko, D. Rus, Tracking a moving object with a binary sensor network, in: ACM SenSys, 2003. [4] N. Bulusu, J. Heidemann, D. Estrin, Adaptive beacon placement, in: ICDCS, 2001. [5] W. Chung, S. Kim, J. Choi, High speed navigation of a mobile robot based on robot’s experiences, in: Proc. of the JSME Annual Conference on Robotics and Mechatronics, 2006. [6] L. Girod, D. Estrin, Robust range estimation using acoustic and multimodal sensing, in: IROS, 2001. [7] T. He, C. Huang, B.M. Blum, J.A. Stankovic, T.F. Abdelzaher, Range-free localization schemes in large scale sensor networks, in: MobiCom, 2003. [8] T. He, R. Stoleru, J.A. Stankovic, Spotlight: low-cost asymmetric localization system fornetworked sensor nodes, in: (demo paper) IPSN, 2005. [9] J. Hightower, R. Want, G. Borriello, SpotON: An indoor 3D location sensing technology based on RF signal strength, UW CSE 00-02-02, University of Washington, Department of Computer Science and Engineering, Seattle, WA, 2000. [10] J. Hill, D. Culler, Mica: a wireless platform for deeply embedded networks, IEEE Micro (2002). [11] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless routing for wireless networks, in: MobiCom, 2000. [12] M. Kushwaha, K. Molnar, J. Sallai, P. Volgyesi, M. Maroti, A. Ledeczi, Sensor node localization using mobile acoustic beacons, in: MASS, 2005. [13] S. Lanzisera, D.T. Lin, K.S.J. Pister, Rf time of flight ranging for wireless sensor network localization, in: WISES, 2006. [14] A. Ledeczi, P. Volgyesi, J. Sallai, B. Kusy, X. Koutsoukos, M. Maroti, Towards precise indoor rf localization, in: HotEmNets, 2008. [15] A. Mainwaring, J. Polastre, R. Szewczyk, D. Culler, Wireless sensor networks for habitat monitoring, in: WSNA, 2002. [16] M. Maroti, et al. Radio interferometric geolocation, in: Sensys, 2005. [17] D. Moore, J. Leonard, D. Rus, S. Teller, Robust distributed network localization with noisy range measurements, in: SenSys, 2004. [18] D. Niculescu, B. Nath, Ad hoc positioning system (APS), in: GLOBECOM, 2001. [19] N.B. Priyantha, H. Balakrishnan, E. Demaine, S. Teller, Mobile-assisted localization in wireless sensor networks, in: INFOCOM, 2005. [20] A. Savvides, C. Han, M.B. Strivastava, Dynamic fine-grained localization in adhoc networks of sensors, in: MobiCom, 2001. [21] A. Savvides, H. Park, M.B. Srivastava, Mobile Networks and Applications (2003). [22] Y. Shang, W. Ruml, Y. Zhang, M.P.J. Fromherz, Localization from mere connectivity, in: MobiHoc, 2003. [23] M. Sichitiu, V. Ramadurai, Localization of wireless sensor networks with a mobile beacon, in: MASS, 2004. [24] G. Simon, M. Marti, A. Ldeczi, G. Balogh, B. Kusy, A. Ndas, G. Pap, J. Sallai, K. Frampton, Sensor network-based countersniper system, in: SenSys, 2004. [25] A. Smith, H. Balakrishnan, M. Goraczko, N.B. Priyantha, Tracking moving devices with the cricket location system, in: MobiSys, 2004. [26] M.M.B. Tariq, M. Ammar, E. Zegura, Message ferry route design for sparse ad hoc networks with mobile nodes, in: MobiHoc, 2006. [27] C. Wang, Y. Ding, L. Xiao, Virtual ruler: mobile beacon based distance measurements for indoor sensor localization, in: MASS, 2006.

656

Y. Ding et al. / J. Parallel Distrib. Comput. 70 (2010) 644–656

[28] K. Whitehouse, The design of calamari: an ad-hoc localization system for sensor networks, Master’s Thesis, University of California at Berkeley, 2002. [29] K. Whitehouse, D. Culler, Calibration as parameter estimation in sensor networks, in: WSNA, 2002. [30] K. Whitehouse, C. Karlof, A. Woo, F. Jiang, D. Culler, The effects of ranging noise on multihop localization: an empirical study, in: IPSN, 2005. [31] Y. Xu, J. Heidemann, D. Estrin, Geography-informed Energy Conservation for Ad Hoc Routing, in: MobiCom, 2001. [32] W. Zhao, M. Ammar, E. Zegura, A message ferrying approach for data delivery in sparse mobile ad hoc networks, in: MobiHoc, 2004.

Yong Ding received the B.S. and M.S. Degrees from Southeast University, China, in 2001 and 2004, respectively. He is currently a Ph.D. candidate in computer science at Michigan State University. His research interests are in the areas of distributed systems and computer networking, including wireless sensor networks, vehicular ad hoc networks, and wireless mesh networks. He is a member of the IEEE.

Chen Wang received the B.S. and M.S. Degrees from Northeastern University, China, and the Ph.D. Degree in computer science and engineering from Michigan State University in 2007. He is currently working at Google. His research interests are in the areas of distributed systems and computer networking.

Li Xiao received the B.S. and M.S. Degrees in computer science from Northwestern Polytechnic University, China, and the Ph.D. Degree in computer science from the College of William and Mary in 2002. She is an associate professor of computer science and engineering at Michigan State University. Her research interests are in the areas of distributed and networking systems, overlay systems and applications, and wireless sensor and mesh networks. She is a member of the IEEE.