Indoor Location Estimation in Sensor Networks using AI Algorithm

J´ozsef D´aniel Dombi

Department of Software Engineering, University of Szeged,

Arpad t´er 2., Szeged, Hungary

Keywords:

Indoor Location and Tracking, Sensor Network, Fingerprinting, Machine Learning.

Abstract:

To determine the indoor location of a person or object, we can use a suitable wireless network. There are

different kinds of wireless networks available for this. Independent of the type of the network, using RSSI

it is possible to ﬁnd the position of the moving person close by. Here, we present Wireless Sensor Network

and apply it in a real environment. We will mainly concentrate on locating a person using standard artiﬁcial

intelligence methods. In our system we deﬁne nodes (the ﬁngerprint), and supervised learning algorithms that

should predict these nodes. In addition, we test whether we can get nice results if we change the granularity of

the nodes. Real simulation demonstrates that this system can supply the current position of the moving person

with good accuracy.

1 INTRODUCTION

Here, a locating system is used for tracking and deﬁn-

ing the current position of a person or object. The

most important distinguishing feature of such a sys-

tem is the type of wireless communication used, and

the application information presented to the user. The

granularity of the position can vary from one applica-

tion to another. For example, ﬁnding out whether a

person is in a room requires less information, while

locating a person who is sitting in front of a desk re-

quires more accurate information.

Therefore, many different systems and technolo-

gies have been proposed. GPS devices are avail-

able for everyday use in modern outdoor applications

(Enge and Misra, 1999). The GPS system has a lim-

ited accuracy, and can be used where satellites are

”visible”, because buildings block the GPS transmis-

sions. The earliest investigation for indoor position-

ing was done by Bahl et al. who observed that an RF

signal source exhibits spatial variation, but is consis-

tent in time. They created a system called Radar (Bahl

and Padmanabhan, 2000). They used four 802.11 ac-

cess points to locate a laptop at its true position to an

accuracy of 2-3 meters. Since then, there have been a

lot of improvements in Radar’s ﬁngerprint matching

algorithms (Agrawala and Shankar, 2003) (Haeberlen

et al., 2004) (Ladd et al., 2005).

These studies showed that the Received Signal

Strength Indicator (RSSI) has a larger variation be-

cause it is subject to the detrimental effects of fading

and shadowing.

Other techniques, such as Active Badge (Hopper

et al., 1993) and a commercial system like Versus

(Versus, 2012), use infrared emitters and detectors

to achieve an accuracy of 5-10m. Active Bet (Har-

ter et al., 1999) (Ward and Jones, 1997) and Cricket

(Priyantha et al., 2000) combine the RF and ultra-

sound signal to estimate the distance. These sys-

tems have accuracies ranging from a few meters to a

few centimeters. In a commercial system (Ubisense,

2012), ultra-wideband emitters and receivers have

been used to realize indoor locations.

In this study we use a wireless sensor network.

If a large number of sensors are deployed, the net-

work can monitor large areas. We can apply a sen-

sor network in a variety of situations like those for

monitoring the environment. Sensor nodes can mea-

sure temperature, a heartbeat, humidity and so on.

However, collecting a large amount of data leads to

an increase in trafﬁc and in the energy consumption

of sensors. Moreover, increasing the data collection

time has a negative impact on the location data col-

lection method. In a wireless sensor network it is vi-

tal to keep the energy consumption low. Our Sensor

Network protocol is similar to the ZigBee (ZigBee,

2012) protocol, which includes IEEE 802.15.4 for

MAC and PHY. Here, we implemented a positional

estimation technique based on standard artiﬁcial in-

telligence methods using RSSI in a sensor network

and evaluated its position-estimation ability. The re-

mainder of this paper is organized as follows. Sec-

349

Dombi J..

Indoor Location Estimation in Sensor Networks using AI Algorithm.

DOI: 10.5220/0004098303490352

In Proceedings of the 14th International Conference on Enterprise Information Systems (ICEIS-2012), pages 349-352

ISBN: 978-989-8565-10-5

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

tion 2 outlines the standard AI methods, then Section

3 describes the experimental setup. After, Section 4

presents the result of our experiments. In the last sec-

tion, we summarize our ﬁndings and draw some per-

tinent conclusions.

2 AI METHODS

In our system the signal strengths are got by a router.

Currently, different routers send the RSSI to the PC.

More than three RSSI values are used to determine

the position of the node inside the building. First, we

have to investigate the relationship between the dis-

tance and signal strength from a given router point. If

one knows the distances from a node to at least three

different routers, one can calculate the position of the

node in the system.

In a real environment the power received is a very

complex function of distance. Even if a good model is

available to determine the position of the node, it still

requires a lengthy calculation. Hence, the location

of the RSSI is more complicated and it is harder to

solve. In our model, we simplify the system. We do

not worry about calculating the exact position of the

object. For us, it is sufﬁcient to determine the nearest

node (ﬁngerprint).

Standard artiﬁcial intelligence methods offer a

good solution for estimating the location and reduc-

ing the distance error. Here, we implemented the de-

cision tree model and neural network model. An ex-

act knowledge of the position is not required by either

method. We can train and use the methods without

asking for it. Both methods have good classiﬁcation

capabilities and are suitable for our purpose, where

we wish to determine the location that best matches

the observed signal strength data.

Using a decision tree means we have to generate

all possible decision trees that correctly classify the

training set and then choose the simplest one. The

number of such trees is ﬁnite, but very large. One

of the most widely used decision tree method is ID3

(Quinlan, 1986). It constructs the simple decision

tree, but this approach cannot guarantee that better

trees have not been overlooked. The basic structure

of ID3 is iterative. The window, which is a subset of a

training set, is chosen at random and a decision tree is

formed from it. ID3 examines all candidate attributes

and chooses attribute A to maximize the gain. This

tree correctly classiﬁes all objects in the window. All

other objects are then classiﬁed using the tree. If the

tree returns the correct answer for all these objects, it

is then correct for the entire training set and the pro-

cess terminates. If not, a selection of the incorrectly

classiﬁed objects is added to the window and the pro-

cess continues. Recent articles (Yim, 2008) have ex-

amined how a decision tree works in a location system

based on a ﬁngerprint, and it is found that the accu-

racy of the decision tree is no worse than a Neural

Network or Bayesian system.

A neural network is capable of representing the

relationship between the inputs (signal strengths) and

outputs (nodes). The learning strategy should calcu-

late the free parameters of the model (also called the

”weights” of the network). Here, the standard multi-

layer perceptron (MLP) is implemented. The archi-

tecture of MLP is organized as follows: the signals

ﬂow sequentially through the different layers from the

input to the output layer. For each neuron, it ﬁrst cal-

culates a scalar product between a vector of weights

and the vector given by the output of the previous

layer. A transfer function is then applied to the result

to produce the input for the next layer. A commonly

applied transfer function is the sigmoid function. In

a single hidden layer, if the number of hidden layers

is sufﬁciently large then any continuous function can

be approximated to some desired accuracy. Roberto

Battiti et al. (Battiti et al., 2002) examined how a neu-

ral network might be used to locate an object. They

found that with MLP it is possible to determine the

position of the person within 1.82 meters.

In our study, we compare the performance of both

methods to see how well they determine the location

of an object in a sensor network environment.

3 EXPERIMENTAL SETUP

Our experimental testbed is located on the ﬁrst ﬂoor

of a 2-storey building. We deﬁne nodes (position of

the ﬁngerprint), and the distances between the nodes

are equal, namely a distance less than 2 m. Part of

the layout of the ﬂoor and position of ﬁngerprint are

shown in Figure 1. In the tests, we employed a special

type of sensor network called RTLS (RTLS, 2012).

We placed four routers per room and two on the cor-

ridor at the locations indicated in Figure 2.

Figure 1: Map of the ﬂoor and position of each ﬁngerprint.

ICEIS2012-14thInternationalConferenceonEnterpriseInformationSystems

350

Figure 2: The position of routers inside the building.

In this ﬁgure, routers are represented by numbers

(i.e.: 211, 240, 243, 168, etc.) and we see that the

position of each router is usually in the corner of a

room. Users wear a transmitter (also called a tag) de-

vice on their wrist as a watch (see Figure 3, which has

a unique ID called the address.

Figure 3: Two different kinds of watch. Both of them func-

tion as a sensor.

The tag can measure and transmit the temperature

and battery level; and, of course, the routers can mea-

sure the RSSI value. It is also possible to send audio

data through this sensor network.

By default, the tag will send a broadcast message

every 4 seconds. When a router gets a message it can

transmit this data to the coordinator (zero in Figure

4, a special router). The packet received by the co-

ordinator contains the address of the measured tag,

the RSSI value measured by the router, and any other

data measured by the tag. In this network there is a

time delay in the routers. The routers wait for a while

to receive RSSI values, then they aggregate them and

transmit this data to the coordinatoras a single packet.

There is a size limit of the packet so in this way the

router should be able to send a packet to coordina-

tor every 400 milliseconds. The aggregated package

contains only the latest RSSI value received from the

tag. As we mentioned above, the coordinator can re-

ceive packets from different routers and it forwards

them to a PC. The program running on this PC can

collect the RSSI values. The primary task of the pro-

gram is to determine which measurement belongs to

the given tag at any one time. Our network is self-

organized. This means that a tag can communicate

with the routers and these routers send the received

information on to the next router, which is closer to

the coordinator and is connected to the PC. A tag tries

to reach the nearest router, and if it cannot commu-

nicate with this router then it will search for another

router. Figure 4 shows how communication is estab-

lished and maintained. In this ﬁgure we can see that

there is coordinator (C1), which is connected to the

PC and there are six routers and a Tag (E1).

Figure 4: Communication between routers and tags.

We generated a ﬁngerprint by performing calibra-

tion measurements. For each node, we measured over

20 values in a second and stored the RSSI values ob-

tained by the router. These were the reference values

that were used for testing the system.

4 RESULTS

As mentioned previously, we tested the AI methods

on a ﬁrst ﬂoor of a building. We positioned the

routers and coordinator. First we had to collect sam-

ples and then we used the cross-validation method.

This method partitions a given data sample into com-

plementary subsets. Then we performed an analysis

on one subset (called the train set), and validated our

analysis on the other subset (called the test set). Mul-

tiple rounds of cross-validation were performed using

different partitions, and the average over the rounds

was the result of the validation.

It should be noted that the objective of our train-

ing algorithm was to build a model with good gener-

alization capabilities when it was tested with values

not present in the train set. The number of parame-

ters and the length of the train phase determined the

goodness of the generalization.

In a real environment it may happen that the given

tag cannot reach the router (missing value). In that

case, we deﬁne the worst RSSI value. In addition we

deﬁne a new attribute that contains this information.

When the value is one, the router receives a signal,

and when the value is zero, the router doesn’t receive

an RSSI value of the given time. The maximum value

of RSSI that we measured was -54dBm and minimum

value was -90 dBm. For each measurement, out of 14

routers 6 on average send a message saying that they

IndoorLocationEstimationinSensorNetworksusingAIAlgorithm

351

receive an RSSI value, and only 3 routers on average

can measure valuable RSSI values - which means that

they can measure values better than -85 dBm. We cre-

ated different kinds of tests which varied the granular-

ity of the nodes: single position, triple position, and

the room. Single position means that we would like

predict the current position of the object. Triple posi-

tion means that we aggregated 3 nearest node values

into one, and we tried to predict this new position.

In this case, we were only interesed in locating the

object in a certain part of the room. Room position

means that we merged all the node values in the room

into a single node in order to locate the object. The

results are shown in the following table.

Table 1: The results of the methods.

Granularity of nodes Decision tree Neural network

Single Position 38% 40%

Triple position 65% 53%

Room 91% 89%

As we see, the two methods have a similar per-

formance in most cases. The percentage value tells

us the degree of certainty of location an object. We

tried different kinds of parameter input for the two

learning methods and we obtained similar results. In

the decision tree, we get the whole tree and examine

the decisions. The decision tree has an average size of

250 and an average number of leaves around 125. The

time needed for the learning method and the evalua-

tion of the values is less for a tree than that for a neural

network.

5 CONCLUSIONS

Many indoor positioning methods have been pub-

lished that can be used in a variety of situations. For

any kind of wireless network, the ﬁngerprint method

is the most commonly used approach. Previous stud-

ies showed that AI algorithms can perform well in lo-

cating an object. These studies used different types

of networks. In this paper we compared two different

kinds of AI method in a wireless sensor environment,

which is similar to the ZigBee network. The band-

width of this network is very low, but it can transmit

audio and data measurements in real time with just

one radio chip. We carried out different kinds of tests

using this wireless sensor network, and we discovered

that in most cases the decision tree and neural network

approaches have a similar performance. When we in-

crease the granualty of the nodes, we get much better

results in terms of accuracy.

ACKNOWLEDGEMENTS

The study presented here was supported by the Hun-

garian national grant GOP-1.1.1-07. I would like to

thank

Akos Kiss for his valuable advice, and also

P´eter Kenderesi, P´eter Moln´ar and Bal´azs Szab´o for

providing position data.

REFERENCES

Agrawala, A. K. and Shankar, A. U. (2003). WLAN loca-

tion determination via clustering and probability dis-

tributions. In PerCom.

Bahl, P. and Padmanabhan, V. N. (2000). RADAR: An in-

building RF-based user location and tracking system.

In INFOCOM.

Battiti, R., Villani, A., Villani, R., and Nhat, T. L. (2002).

Neural network models for intelligent networks: De-

riving the location from signal patterns. In Citeseer.

Enge, P. and Misra, P. (1999). Special issue on gps: The

global positioning system. In Proceedings of the of

the IEEE.

Haeberlen, A., Flannery, E., Ladd, A. M., Rudys, A., Wal-

lach, D. S., and Kavraki, L. E. (2004). Practical

robust localization over large-scale 802.11 wireless

networks. In Proceedings of the 10th Annual Inter-

national Conference on Mobile Computing and Net-

working.

Harter, A., Hopper, A., Steggles, P., Ward, A., and Webster,

P. (1999). The anatomy of a context-aware applica-

tion. In Proceedings of the Fifth Annual ACM/IEEE

International Conference on Mobile Computing and

Networking (MobiCom-99).

Hopper, A., Harter, A., and Blackie, T. (1993). The active

badge system. In INTERCHI’93 Conference on Hu-

man Factors in Computing Systems.

Ladd, A. M., Bekris, K. E., Rudys, A., Kavraki, L. E., and

Wallach, D. S. (2005). Robotics-based location sens-

ing using wireless ethernet. Wireless Networks.

Priyantha, N. B., Chakraborty, A., and Balakrishnan, H.

(2000). The cricket location-support system. In MO-

BICOM.

Quinlan, R. (1986). Induction of decision trees. Machine

Learning, 1.

RTLS (2012). Rtls. http://www.rtls.eu.

Ubisense (2012). Ubisense. http://www.ubisense.net.

Versus (2012). Versus. http://www.versustech.com.

Ward, A. and Jones, A. (1997). A new location technique

for the active ofﬁce. In IEEE Personal Communica-

tions.

Yim, J. (2008). Introducing a decision tree-based indoor

positioning technique. Expert Syst. Appl, 34(2).

ZigBee (2012). Zigbee. http://www.zigbee.org/Speciﬁcati

ons.aspx.

ICEIS2012-14thInternationalConferenceonEnterpriseInformationSystems

352