PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY

METHOD FOR WIRELESS SENSOR NETWORK AD-HOC PERSON

LOCALIZATION

Tobias G

adeke, Johannes Schmid, Wilhelm Stork and Klaus D. M

uller-Glaser

Institute for Information Processing Technologies (ITIV), Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany

Keywords:

Localization, Wireless, Sensor Networks, Person, Pedestrian.

Abstract:

The problem of localization and navigation in areas without any or with only limited access to global naviga-

tion satellite systems (GNSS) is still not solved. This is especially the case for person localization applications

as persons tend to spend a good part of their time in buildings or in cities (urban canyons). One possibility

to approach this issue is to use wireless sensor network (WSN) technology. Especially scenarios that require

ad-hoc person localization like ﬁreﬁghters that enter a burning building or similar setups, WSN seem to be

a promising solution. However, if the node density is low or if the scenario also requires localization in un-

covered areas, an additional localization method is required. Pedestrian dead reckoning (PDR) is an ideal

complementary method to achieve short term accurate localization under these assumptions.

In this paper, an approach to PDR with low processing power for the use in WSN with a hip mounted inertial

measurement unit (IMU) is presented. The purpose of the system is to provide a localization and tracking

solution if temporarily none or only few anchor nodes are within communication range. This is achieved by

detecting steps, estimating the length of each step and determining the step direction in WSN coordinates.

We experimentally evaluate the system under varying environmental conditions and show that the concept is

a promising solution for the intended applications.

1 INTRODUCTION

To solve the challenge of node localization, a lot of

research has been undergone in the ﬁeld of localiza-

tion in wireless sensor networks (WSN) during the

last few years. The problem of correlating a measured

value with its time and its location is a fundamental

requirement in most applications. Figure 1 shows one

example of an application of such WSN for localiz-

ing and tracking persons in scenarios where global

navigation satellite systems (GNSS) are not available.

Other possible application areas of a system that can

provide localization under such conditions are for ex-

ample ﬁreﬁghters or other rescue forces that want to

logistically coordinate a mission. In these scenar-

ios an easily installable and robust solution to the lo-

calization problem is required. To achieve this, one

way could be to use a dedicated WSN for localization

with the additional feature of providing communica-

tion within the network. Additionally, this allows the

integration of other sensor readings, e.g., motion or

movement sensors, or smoke detectors, which leads

to a modular extensible localization and surveillance

Figure 1: Application scenario: Localization of rescue

forces in and around destroyed buildings (G

adeke et al.,

2011b).

system (G

adeke et al., 2011b). Other application ex-

amples with similar requirements are security forces

or localization of builders, tools and measurement

equipment on construction sites. Another use case

can be found in hospitals where patients and doctors

139

Gädeke T., Schmid J., Stork W. and D. Mueller-Glaser K..

PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY METHOD FOR WIRELESS SENSOR NETWORK AD-HOC PERSON LOCALIZATION.

DOI: 10.5220/0003804801390148

In Proceedings of the 1st International Conference on Sensor Networks (SENSORNETS-2012), pages 139-148

ISBN: 978-989-8565-01-3

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

need to be localized for navigation purposes and opti-

mized scheduling.

In these intended applications, required accuracies

are in the range of a few meters, i.e. on room level.

The simplest method to achieve this in WSN is by

evaluating the received signal strength (RSS) of ra-

dio packets from anchor nodes with known positions

(Schmid et al., 2011a).

Especially in indoor scenarios however, location

estimates from RSS range-based approaches are of-

ten subject to ﬂuctuations due to the badly predictable

behavior of RSS values. To cope with this, stochastic

ﬁltering can be used and a system model of the mov-

ing person can be incorporated. Also, inertial sensors

can be used to improve the localization accuracy to

a certain degree (Schmid et al., 2011b). However, in

cases where the area of interest cannot be fully cov-

ered by the WSN or the anchor node density is too

low, a complementary localization system is needed

to provide short-time accurate localization during the

periods without WSN coverage.

In this paper short-time accuracy is achieved by

integrating a pedestrian dead reckoning (PDR) unit

into a WSN. Because of the low processing capabili-

ties of the sensor node’s microcontroller unit (MCU)

the application requires an approach which does not

include complex algorithms. A hip mounted inertial

measurement unit (IMU) is considered for step de-

tection, step length estimation and attitude calcula-

tion. We present results of a comprehensive experi-

mental evaluation of this concept under different en-

vironmental conditions which supplement ﬁrst results

presented in (G

adeke et al., 2011a).

The remainder of this paper is structured as fol-

lows: In Section 2 a short survey of the state of the art

and the related work in the ﬁeld is given. In Section 3

the developed concept and information processing is

introduced. Some details on the implementation and

the undergone experimental study are given in Section

4 and the results are presented and discussed in 5. The

paper is concluded in Section 6 and an overview of the

next steps is given in 7.

2 STATE OF THE ART

The last years led to a price decline in micro electro

mechanical systems (MEMS) due to the technologi-

cal improvements for these components. Along with

this development also the interest in pedestrian navi-

gation solutions increased and many research groups

based their work on MEMS IMU inertial localiza-

tion and navigation. There are two main approaches

for placing the IMU on the human body. Often con-

sidered is the placement of the IMU on the foot of

the user. Other placements on the body need a step

counting approach for localization which will be ex-

plained at the end of this paragraph. With a foot

mounted IMU strapdown algorithms which are well

known from aerospace or marine navigation can be

applied. Usually high precision inertial sensors are

used and allow for a small error growth. The atti-

tude is obtained by a mathematical integration of the

turn rates and the position is calculated from a dou-

ble integration of the acceleration signal (Titterton,

2004). Today’s MEMS inertial sensors still suffer

from a large drift. To compensate for these drifts dif-

ferent additional system inputs have been proposed.

The zero-velocity update (ZUPT) is among the most

famous and sets the velocity to zero during each

stance phase of the gait cycle (Foxlin, 2005), (Jimenez

et al., 2010). Angular rate updates might be fused

by the zero-angular rate update (ZARU) analogously

(Jimenez et al., 2010). More recently, a heuristic drift

reduction (HDR) algorithm which makes the assump-

tion that pedestrians are mostly walking along straight

paths has been used in different peculiarities (Boren-

stein et al., 2009), (Jimnez et al., 2011). Typically,

slightly bended paths are the major challenge to these

algorithms but it works very well for long corridors or

streets. The results obtained by all approaches with

foot-mounted IMUs are very competitive. The error

for these systems are typically on the order of only a

few percent of the distance traveled (Wan and Foxlin,

2010). But, achievable accuracies depend strongly

on the inertial sensors performance and the applied

data processing. However, behind these systems a

complex system model is used and hence in a state

space model, a large state vector is needed. Also, the

dynamics occurring on the foot are quit high. That

makes it necessary to use a high update rate for the

data processing. Both of these characteristics require

fast processing capabilities which also imply shorter

battery lifetimes or larger batteries. For integration in

a WSN this is contradictory to the low power hard-

ware used within such a system.

On the other hand, beside the IMU placement on

the foot many other positions on the human body

have been evaluated. (Beauregard, 2006) consid-

ers the head, rucksack structures are used in (Nie-

dermeier et al., 2009) and another popular position

is the hip (Kourogi and Kurata, 2003). For all of

these approaches strapdown algorithms with ZUPT

and ZARU cannot be used. There is no possibility to

correct the system model and hence, the error evolves

too fast with current MEMS inertial sensors. For

these IMU positions step recognition together with

step length estimation is usually considered. In terms

SENSORNETS 2012 - International Conference on Sensor Networks

140

of accuracy, the performance for these approaches

cannot compete with the performance of the foot-

mounted systems. Relative errors are typically in the

range of 10 % of traveled distance (Randell et al.,

2003). Further improvements can be achieved by

adapting parameters and algorithms to the user’s cur-

rent posture (Sun et al., 2009). However, different

properties of the sensor signal are often requires for

such an activity classiﬁcation. An example is the

frequency domain representation (Fourier transforma-

tion, e.g. FFT calculation) but also all kind of digital

ﬁlters. These algorithms are based on matrix com-

putations which are typically done by a digital sig-

nal processor (DSP). Especially with high data rates

and ﬂoating point operations these are considered too

complex for a typical MCU used on a wireless node.

To achieve pedestrian localization with ultra-low

power hardware used in WSN a hip mounted IMU

is considered. Beside the lower power consumption

this placement is also more practical and realistic for

the applications considered in this paper. The lower

dynamics in the acceleration signal also allows for

a lower update rate which further reduces the com-

putational complexity. Furthermore hip based ap-

proaches are easily portable to off-the-shelf smart

phones (Gusenbauer et al., 2010), (Serra et al., 2010).

This opens a wide range of consumer applications

like navigation in exhibition halls or airports, but also

seamless indoor/outdoor navigation in underground

parking and shopping centers.

3 SYSTEM CONCEPT

3.1 Pedestrian Dead Reckoning

In this paper an IMU of the model MTi-G from Xsens

Technologies B.V. was used. The IMU incorporates

three-dimensional acceleration sensors, gyroscopes

and magnetometers. An external GPS antenna can

be connected to further improve the internal heading

estimation algorithm. Also, when the GPS connec-

tion is used a pressure sensor stabilizes the height es-

timation. The IMU features an internal DSP which

outputs the attitude based on a preconﬁgured preset.

These presets inﬂuence the parameters of the attitude

calculation. For example, it can be selected that sen-

sor readings from the magnetometers should be incor-

porated in the calculation or not. The outputs of the

IMU are calibrated sensor readings from all sensors

and also the internally calculated attitude based on the

chosen preset. This is a useful feature for the consid-

ered system because the sensor nodes MCU does not

have to handle the complex calculation. A drawback

Figure 2: Acceleration energy and step detection threshold.

of this approach is that the algorithms on the sensor

node have to rely on the attitude and no further correc-

tions (like HDR, tuning of sensor readings depending

on the application) on the attitude can be applied.

As stated in Section 2 already, current MEMS IMU

do not provide the accuracy needed for strapdown al-

gorithms when it is placed somewhere on the human

body. Our approach to determine the user’s position

is based on step recognition which updates the posi-

tion estimation at each detected step. Therefore, it is

also necessary to know direction and length of each

step. These three issues will be covered in the fol-

lowing sections after a short clariﬁcation on the used

coordinate frames has been introduced.

3.1.1 Coordinate Frames

The resulting position estimation is with reference to

the navigation frame (n-frame) which represents a lo-

cal tangent plane (LTP) on the earth’s surface. How-

ever, sensor measurements from the IMU are in body

frame (b-frame)-coordinates of the IMU. To quantify

the pedestrian’s step direction an additional coordi-

nate frame is introduced. This is referred to as hu-

man frame (h-frame) in which the x-axis describes the

forward-walking direction, the y-axis is perpendicular

in the horizontal plane (sidewards) and the z-axis is

perpendicular to both other axes and roughly aligned

with the human body (up-direction).

3.1.2 Step Recognition

The detection of successive steps is based on a peak

and threshold analysis. The acceleration energy E(~a

)

is calculated in every time step k from all three dimen-

sions of the acceleration signal ~a

E(~a

) = a

x,k

+ a

y,k

+ a

z,k

(1)

Figure 2 shows the acceleration energy pattern with

the threshold applied. Additionally, multiple peaks

within a time-frame of 300 ms after a successfully de-

tected step are rejected. This has a similar effect as

ﬁltering the signal with a low pass ﬁlter beforehand

and prevents multiple detections of the same step.

The threshold is set to a value of 10 (m/s

)

which

applies for slow walking patterns, but also holds for

faster walking and running.

PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY METHOD FOR WIRELESS SENSOR NETWORK

AD-HOC PERSON LOCALIZATION

141

3.1.3 Steplength Estimation

For an on-line estimation of the users step length there

exist a number of different approaches. However, they

all perform very similarly (Jahn et al., 2010). Some

approaches need a calibration for each particular user

which is not desired for ad-hoc localization (Zhao,

2010). For the purpose of this paper the Weinberg Al-

gorithm is chosen because of its simplicity and miss-

ing calibration procedure (Weinberg, 2002). The step

length SL is calculated from the maximum and min-

imum measured acceleration a in z-direction of the

n-frame. A constant parameter W can be set for unit

transformations.

SL =

max(a

) − min(a

)·W (2)

3.1.4 Step Direction

The step direction, i.e. the alignment of the IMU to

the human body, needs to be known in coordinates

of the b-frame so that it can be transformed to the

n-frame. Assuming the user is walking forward, the

step direction is given in the h-frame by the vector

[1,0,0]

. For normal and fast walking the step direc-

tion is characterized by a larger acceleration in move-

ment direction than the sidewards acceleration. It can

be obtained by a principal component analysis (PCA)

over these two components of acceleration (Kourogi

and Kurata, 2003). From the raw-acceleration data

in the b-frame the gravitational component g and the

acceleration a

z,k

has to be removed. These can be ob-

tained with the direction cosine matrix (DCM) C

directly.

= C

· [0,0,g]

(3)

= C

· [0,0,a

z,k

]

(4)

Applying the PCA to the remaining two components

of acceleration yields the step direction

sd in b-frame

coordinates.

= PCA(~a

− g

− a

) (5)

It can, however, not yet been distinguished between

forward and backward direction. Similar to (Kourogi

and Kurata, 2003) the alternation of maximums in x-

and z-components of the h-frame acceleration is used

to determine the step direction completely. With this

kind of step direction estimation, also the relationship,

i.e., the transformation matrix between the b- and h-

frame is given by the three principal component vec-

tors returned by the PCA.

Because calculating the step direction puts a heavy

workload on the MCU, the procedure is done for ini-

tialization only. Also, the PCA delivers unreliable re-

sults if the person to localize is running or moving

Figure 3: Overview of the measurements and data process-

ing.

slowly. For these walking patterns accelerations in

other than the walking direction become more domi-

nant. This results in a noisy step direction which has

a great inﬂuence on the position estimation.

For the intended use cases it is sufﬁcient to call the

procedure once after the device is attached to the user.

For other applications (e.g. smart phones), the for-

ward direction can change if the device is taken out

of the pocket and being put back in. This implies a

recalibration procedure whenever a reliable walking

pattern is recognized.

3.2 System and Measurement Model

To model the system and the measurement inputs, a

modiﬁed version from (Schmid et al., 2011b) is used.

Incoming measurements are processed by means of

a Kalman ﬁlter. Figure 3 shows an overview of the

considered measurement inputs and information pro-

cessing. In contrast to (Schmid et al., 2011b) PDR po-

sition updates are considered as an external input ~u

to the system and hence are part of the system model.

ˆx

= A · ˆx

+ B · ˆu

+ ˆw

(6)

Position updates from the WSN are considered as

measurement inputs and are processed in the update

step of the Kalman ﬁlter. For localization without an

IMU a system model which assumes a constant speed

between each time-step is used. Therefore, the state

vector ˆx

models position and velocity. To prevent

drifting away very fast when no sensor readings are

present at all a velocity prediction factor smaller than

1 is chosen in the system model. This is established

by setting the corresponding elements of the system

matrix A to a value smaller than 1, i.e., 0.99.

Fusing PDR position updates with the system

model increases the covariance matrix of the overall

position estimation which represents the typical be-

havior of any PDR system.

SENSORNETS 2012 - International Conference on Sensor Networks

142

Figure 4: IMU sensor node combination (Xsens MTi-G and

ITIV LocNode).

Additionally, to achieve long term stability RSS

measurements to anchor nodes (with known posi-

tions) are used upon availability. As soon as the on-

body node receives a broadcast message from an an-

chor node, distance information is deduced from the

RSS value of the received packet. This information

is fused to the current position estimate in the ﬁl-

ter step of the Kalman ﬁlter. The relationship be-

tween distances and corresponding RSS values has

been evaluated experimentally and is modeled by the

log distance path-loss model (Schmid et al., 2011a).

As these measurements are non-linear an extended

Kalman ﬁlter is used for linearization. Complemen-

tary to the PDR covariance model feeding of RSS po-

sition estimations decreases the covariance matrix of

the system.

Localization based on PDR alone comes to use

whenever no or too few anchor nodes are available.

For example if the person to localize leaves the WSN-

covered area or arrives at a place where the anchor

node density is low. In such situations the PDR sys-

tem has to overtake until the person once again re-

ceives radio packets and the position can be corrected.

4 IMPLEMENTATION AND

EXPERIMENTAL SETUP

The presented evaluation is based on real data but

carried out ofﬂine. This method allows to evaluate

various parameter settings of data processing con-

cepts without having to re-run each experiment sev-

eral times and modify the MCU’s source code. On

the hardware side, a PDR unit to be carried on the

hip of a person to be localized has been developed.

The sensor nodes are also based on our own design.

The sensor network software is based on a proprietary

ZigBee stack implementation. The developed system

was used to record data in various environments in-

cluding in- and outdoor scenarios.

4.1 Hardware

Figure 4 shows the developed PDR-unit consisting of

the Xsens MTi-G IMU and an ITIV LocNode sensor

node(Schmid et al., 2011a). The casing is designed to

ﬁt into a standard camera bag which can be attached

to a belt and carried on the hip of a test person. This

PDR unit allows storing the RSS values of received

radio packets as well as the IMU’s attitude estimation

and acceleration data on an SD-card.

The incorporated Xsens MTi-G IMU provides cal-

ibrated data from acceleration, gyro and magnetic

ﬁeld sensors and also includes a DSP to fuse this data

for attitude estimation. The DSP’s data processing

can be tuned with predeﬁned proﬁles according to the

scenario in which the unit is used. For the under-

gone experimental study, the ”aerospace” proﬁle was

used. This proﬁle results in an inclusion of data from

the magnetic ﬁeld sensors into the attitude estimation

for long term heading stabilization. Although mag-

netic ﬁeld sensors can be disturbed in indoor environ-

ments, they are the only possibility to obtain reliable

long term stability. Because fusing turn rates from

the gyro sensors with magnetic sensor readings, these

magnetic ﬁeld disturbances can be partially corrected.

Also, the IMU’s incorporated GPS receiver and the

barometer were disabled and not used for this work.

4.2 Software

A ZigBee framework is used to setup a WSN with

self-organization and multi-hop capabilities. Local-

ization functionality is implemented in this frame-

work. For the purpose of an off-line evaluation,

the framework also allows to store all received radio

packets on the sensor nodes’ SD-card. Each anchor

node is conﬁgured to broadcast its own position reg-

Table 1: Overview of lengths of the experiments.

Institute

Under-

ground

Football

ﬁeld

Parking

lot

mean

length [m]

700 450 750 1320

mean

length [s]

560 290 480 900

max.

length [m]

1120 510 910 1920

min.

length [m]

390 370 570 1000

PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY METHOD FOR WIRELESS SENSOR NETWORK

AD-HOC PERSON LOCALIZATION

143

(a) Underground car park. (b) Football ﬁeld.

Figure 5: Surroundings, ground truth and reconstructed trajectories for exemplary runs in the 4 experimental setups.

ularly at a rate of 4 Hz. The on-body node processes

acceleration and attitude data at a rate of 10 Hz and

radio packets upon availability.

4.3 Experimental Setup

For a quantitative experimental evaluation of the pro-

posed hybrid PDR WSN localization approach, a lot

of data had to be collected. The goal of the undergone

experimental campaign was to prevent an over-ﬁtting

of the proposed fusion approach to a speciﬁc envi-

ronment. For this purpose, four experiments in the

institute ofﬁce building, an underground car park, a

football ﬁeld and a parking lot were carried out. In

each experiment, several runs of a couple of minutes

were conducted. Figure 5 shows an overview and an

example trajectory of one run in each environment.

A WSN with 62 sensor nodes was deployed in each

environment. To limit the inﬂuences of too many

parameters, one person carrying the IMU-equipped

sensor node was walking more than 20 km in all ex-

periments and corresponding data from the IMU and

WSN was collected. In each experiment multiple runs

with a length between 370 m and 1920 m were per-

formed. The mean lengths in meters and seconds for

each experiment are given in table 1. It can be seen

that the experiments differ in these characteristics.

That means, that different walking speeds and differ-

ent trajectories have been evaluated. Additionally to

the experiments presented in this paper the PDR con-

cept was also evaluated experimentally with different

users, but with shorter runs and without WSN data

available. A comprehensive analysis of different users

walking patterns are part of our current research.

To allow for an evaluation, a reference (ground

SENSORNETS 2012 - International Conference on Sensor Networks

144

Figure 6: Stand-Alone PDR Error evaluation for all runs in

each experiment.

truth) was recorded with a differential GPS (DGPS)

in the outdoor experiments. Indoors, the user walked

various predeﬁned trajectories. A sequence of way

points was set up and the time between reaching each

of them was measured. For every segment between

the way points a constant speed was assumed. This

method of predeﬁning a trajectory cannot guaranty

the quality of a GPS recorded reference trajectory but

is still an adequate way to obtain a ground truth. For

the purpose of this paper a constant IMU alignment

on the test person’s body (step direction) has been as-

sumed for all runs in all experiments. To allow for

a comparison between in- and outdoor experiments

they are set up in 2D only. However, the calculations

are done in 3D and the height of the sensors can be

easily tracked by barometric sensors.

5 DATA EVALUATION AND

DISCUSSION

The gathered data were evaluated off-line with the

methods described in Section 3. The evaluation con-

centrates on the effect of a short-time accurate lo-

calization system (PDR) to bridge time intervals in

which no other localization system is available. At

ﬁrst, the developed stand-alone PDR solution is eval-

uated and the resulting errors are analyzed. After that

it is shown how this PDR system can be used to sta-

bilize a WSN localization system if it is assumed that

the person to be localized leaves the WSN or the WSN

breaks down. When the person enters the WSN again,

the position errors can be corrected.

5.1 Stand-alone PDR Error Analysis

For the developed PDR system, the following error

sources are dominant. From ﬁgure 5 (circle line) it

can be seen that heading estimation indoors is not as

reliable as in the outdoor scenarios. This error is ex-

pected to be introduced by magnetic disturbances in

the indoor environments. As heading is calculated in-

ternally on the Xsens MTi-G’s DSP it cannot be cor-

rected and parameters of the ﬁlter cannot be adjusted.

Alternatively heading could be calculated from the

raw measurements from the IMU, but this would re-

quire a fast processor on the sensor node. To allow

for an implementation on a widely used WSN plat-

form and achieve long battery lifetimes this option is

not considered in this paper.

The other errors are introduced by the calculations

described in Section 3. Each falsely detected step in-

troduces an error of one step length, although the step

recognition is very robust. A more inﬂuencing error

is the step length calculation which adds a small dis-

tance during each step. Assuming a wrong but con-

stant step direction, results in a heading offset which

turns the whole trajectory in n-frame coordinates by a

certain angle around the starting point. This error can

be observed in ﬁgure 5 (c) in the institute experiment

and is the most critical part in the system. Instead of

analyzing each of the error sources alone, a general

analysis of the system’s performance in resulting po-

sition accuracies is undergone.

To quantify the errors of the PDR system, the rel-

ative deviation to the traveled distance of the refer-

ence trajectory is taken into account. Therefore the

PDR system is analyzed without the WSN localiza-

tion. The errors in every run of each experiment are

summed up and the overall relative error for each ex-

periment is calculated. Figure 6 shows the mean er-

ror and standard deviation for all runs of each experi-

ment. The mean error is between 3.5 % and 10 % of

traveled distance. The relative error is characterized

by very high values in the beginning of each run as

the total traveled distance is short and small displace-

ments result in high error quantities. At the end of

each run the relative error typically stabilizes around

5 % of traveled distance. The highest error and stan-

dard deviation is found in the underground parking

experiment. This might be due to higher disturbances

of the magnetic ﬁeld and the resulting attitude estima-

tion difﬁculties.

However, if the developed hip-mounted PDR unit

is solely considered, the presented method allows for

error rates on the order of a few percent of traveled

distance.

PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY METHOD FOR WIRELESS SENSOR NETWORK

AD-HOC PERSON LOCALIZATION

145

(a) Total error in meters. (b) Error in percent of traveled distance.

Figure 7: Error over time for exemplary run.

5.2 Unavailability of WSN Localization

If long term accurate reference data from a WSN is

available, the absolute error remains more or less con-

stant and the relative error converges to zero with op-

eration time of the system. Person localization in

environments where a dense WSN is deployed al-

lows for room-level accuracies (Schmid et al., 2011a).

However, in realistic scenarios it cannot be assumed

that the person to localize always remains in the area

that is covered by the WSN. Also, it has to be con-

sidered that the WSN could fail during the operation.

In such situations, localization can be continued by

means of dead reckoning.

To evaluate this concept the hybrid solution as de-

scribed in section 3 is considered. Then, it is assumed

that the WSN fails at t

down,start

= 0.3 ·t

run

and recov-

ers at t

down,end

= 0.8 ·t

run

in each run with length t

run

During this time, no RSS packets are received and the

IMU is used for PDR alone.

Figure 7 shows how the error evolves over time

during one example run from the experiments. It can

be seen that the total error of the position estimation

increases during the run if only IMU data are con-

sidered for the localization (circle line). Naturally,

the error remains more or less constant on a relative

scale and the downtime of the WSN does not affect

its accuracy as these data are not used. The cross

line shows localization based on data from the WSN

alone. When the network fails at t ≈ 200 s the posi-

tion sticks close to the coordinates where the last mea-

surements were received. However, the considered

position-velocity model (Section 3) draws the posi-

tion estimation away. The mentioned velocity predic-

tion factor in the system model lets the position con-

verge.

On an absolute scale this results in an increasing

error which can also be observed in the trajectories

in ﬁgure 5. As the estimated position stays roughly at

the last known position the error becomes larger when

the person moves further away from that spot. In the

presented scenarios the test person did not move too

far away from the test ﬁeld so the error stays within

the bounds of the experiment’s area. Especially be-

cause the trajectories come full circle in most cases

the errors become smaller again after a certain point

in time. This implies a maximum value for the total

error. If the developed hybrid information process-

ing approach is used, the PDR system allows continu-

ous localization as long as no WSN data are available

and bridges the gap until the WSN covered area is re-

entered (triangle line). This behavior can be observed

in the trajectories in ﬁgure 5 and in the error analysis

in ﬁgure 7. Compared to the reconstructed trajectories

from PDR (circle line) or WSN (cross line) alone the

hybrid information processing approach allows for a

much better localization accuracy if an outage of the

WSN is assumed (triangle line).

5.3 Robustness Issues and Evaluation

Figure 8 compares the resulting mean errors based on

all collected data from the presented experiments. It

shows the mean errors for failing WSN-only local-

ization and the hybrid solution with and without the

WSN failing.

With the WSN failing it can be seen that the WSN-

only solution cannot provide sufﬁcient localization

accuracy anymore. Compared to that, for the hybrid

solution, the resulting localization accuracies in ﬁgure

8 (a) are in the range of a few meters depending on the

surroundings. If the WSN is not available anymore

SENSORNETS 2012 - International Conference on Sensor Networks

146

(a) Error in meters. (b) Error in percent of traveled distance.

Figure 8: Mean error and standard deviation over all runs in all experiments.

the mean errors increase but still allows for reasonable

localization accuracy. The much higher errors in the

parking lot experiment are due to very long runs com-

pared to the other experiments and following longer

times without WSN localization. To further quantify

the localization accuracies the relative errors are taken

into account. Therefore, relative errors during the un-

availability of each run are given in ﬁgure 8 (b). It can

be seen that the relative errors of the parking lot exper-

iment are comparable to the other experiments. In this

analysis the underground parking experiment exhibits

higher values which are introduced by very high rela-

tive errors in the beginning of each run. Because the

underground parking experiment is characterized by

a series of relatively short runs these higher errors in

the beginning contribute to the high values shown in

ﬁgure 8 (b). The high relative errors in the beginning

are introduced because a very short distance has yet

been covered but the initial absolute error is compara-

bly high. Theoretically this results in inﬁnite large er-

rors. Compared to the stand-alone error analysis from

section 5.1 it can be seen that the relative errors are

higher in the analysis with a failing WSN. As men-

tioned before, relative errors in the beginning of the

WSN outage are very high and thus affect the over-

all error. Additionally, the covered distance is shorter

compared to the stand-alone analysis which also has

an impact on the mean relative error.

In general, the PDR approach delivers a vital com-

plementary solution to localization in ad-hoc WSN

scenarios.

6 CONCLUSIONS

This paper evaluated a person localization approach

for WSN by combining RSS localization with PDR.

A hip-mounted IMU, integrated with a wireless

sensor node is used for step detection and step length

estimation of the user. Additionally, the orientation of

the IMU in relation to the body of the person is es-

timated by means of a PCA. The paper evaluates the

system under different environmental conditions and

presents the performance of the system also when the

sensor network is left for a while.

The focus of the evaluated implementation lies

on algorithms which can be carried out on a typical

WSN MCU. Relative errors are on the order of 5 % to

10 % of traveled distance. This is comparable to state

of the art approaches, but considering the low com-

putational complexity it allows for WSN integration

and longer battery lifetimes. the performance does

not allow for long term navigation without any stabi-

lization by an absolute positioning system. But, the

method presented carries the potential for localiza-

tion and navigation applications where generally an

absolute positioning system is present but not very

reliable at all times. PDR guaranties short-term ac-

curacy, while WSN measurements provide long-term

stability. It has been shown that the presented PDR

approach allows to bridge outages of the WSN or also

helps the user to navigate through uncovered areas for

some time. Possible applications for the presented

person localization approach include ﬁreﬁghters or

security scenarios but also construction sites and hos-

pitals. Also, the principal portability on many smart

phones opens a wide range of consumer applications

like navigation in large buildings (e.g. airports, exhi-

bition halls, shopping centers or others). The contri-

bution of this paper is an efﬁcient approach to PDR

with low processing capabilities and a comprehensive

experimental analysis of the system under varying en-

vironmental conditions.

PEDESTRIAN DEAD RECKONING AS A COMPLEMENTARY METHOD FOR WIRELESS SENSOR NETWORK

AD-HOC PERSON LOCALIZATION

147

7 FUTURE WORK

For the next future, it is planned to further study the

attitude estimation done by the integrated DSP inside

the IMU. Therefore, we are currently working on a

new design for a hip-mounted PDR unit. Further-

more, in terms of algorithms design, we are work-

ing on a classiﬁcation of typical movement patterns to

improve and optimize the performance for a speciﬁc

application scenario. Based on the movement classi-

ﬁcation the alignment procedure of the IMU will be

further improved. Also, the initialization process of

the anchor nodes and cooperative approaches will be

evaluated further.

ACKNOWLEDGEMENTS

This work was carried out within the Research

Training Group GRK 1194 ”Self-organizing Sensor-

Actuator Networks” and is supported by the German

Research Foundation (DFG). We would like to thank

the DFG for supporting our work.

REFERENCES

Beauregard, S. (2006). A helmet-mounted pedestrian dead

reckoning system. In Proceedings of the 3rd In-

ternational Forum on Applied Wearable Computing

(IFAWC 2006).

Borenstein, J., Ojeda, L., and Kwanmuang, S. (2009).

Heuristic reduction of gyro drift for personnel track-

ing systems. Journal of Navigation, 62:41–58.

Foxlin, E. (2005). Pedestrian tracking with shoe-mounted

inertial sensors. IEEE Comput. Graph. Appl.,

25(6):38–46.

adeke, T., Schmid, J., Stork, W., and M

uller-Glaser, K. D.

(2011a). Pedestrian dead reckoning for person local-

ization in a wireless sensor network. In Proceedings of

International Conference on Indoor Positioning and

Indoor Navigation (IPIN).

adeke, T., Schmid, J., Stork, W., Reinhardt, M., Markus,

M., and Blank, T. (2011b). Ad-hoc sensor network for

building monitoring and person localization. In Pro-

ceedings of the 1st International Conference on Wire-

less and Mobile Networks (ICWMCN).

Gusenbauer, D., Isert, C., and Kroesche, J. (2010). Self-

contained indoor positioning on off-the-shelf mobile

devices. In Proc. Int Indoor Positioning and Indoor

Navigation (IPIN) Conference, pages 1–9.

Jahn, J., Batzer, U., Seitz, J., Patino-Studencka, L., and

Gutirrez Boronat, J. (2010). Comparison and evalu-

ation of acceleration based step length estimators for

handheld devices. In Proc. Int Indoor Positioning and

Indoor Navigation (IPIN) Conference, pages 1–6.

Jimenez, A. R., Seco, F., Prieto, C., and Guevara, J. (2010).

Indoor pedestrian navigation using an ins/ekf frame-

work for yaw drift reduction and a foot-mounted imu.

In WPNC 2010: 7th Workshop on Positioning, Navi-

gation and Communication 2010.

Jimnez, A., Seco, F., Zampella, F., Prieto, C., and Guevara,

J. (2011). Improved heuristic drift elimination (ihde)

for pedestrian navigation in complex buildings. In In-

ternational conference on indoor Positioning and In-

door Navigation - IPIN.

Kourogi, M. and Kurata, T. (2003). Personal position-

ing based on walking locomotion analysis with self-

contained sensors and a wearable camera. In Proc.

Second IEEE and ACM Int Mixed and Augmented Re-

ality Symp, pages 103–112.

Niedermeier, H., Ameres, G., Pany, T., and Eissfeller, B.

(2009). First results from supporting long coherent

cdma correlations by a mems ins and a pedestrian nav-

igation system for hs-gnss applications. In Proc. 6th

Workshop Positioning, Navigation and Communica-

tion WPNC 2009, pages 5–13.

Randell, C., Djiallis, C., and Muller, H. (2003). Personal

position measurement using dead reckoning. In ISWC

’03: Proceedings of the 7th IEEE International Sym-

posium on Wearable Computers, page 166, Washing-

ton, DC, USA. IEEE Computer Society.

Schmid, J., Beutler, F., Noack, B., Hanebeck, U., and Mller-

Glaser, K. (2011a). An experimental evaluation of

position estimation methods for person localization in

wireless sensor networks. Wireless Sensor Networks,

pages 147–162.

Schmid, J., Gaedeke, T., Stork, W., and Mueller-Glaser,

K. D. (2011b). On the fusion of inertial data for signal

strength localization. In Proceedings of the 8th Work-

shop on Positioning, Navigation and Communication.

Serra, A., Dess

ı, T., Carboni, D., Popescu, V., and Atzori,

L. (2010). Inertial navigation systems for user-centric

indoor applications.

Sun, Z., Mao, X., Tian, W., and Zhang, X. (2009). Activity

classiﬁcation and dead reckoning for pedestrian nav-

igation with wearable sensors. Measurement Science

and Technology, 20(1).

Titterton, David H. ; Weston, J. L. (2004). Strapdown in-

ertial navigation technology. Progress in astronautics

and aeronautics ; 207. Institution of Electrical Engi-

neers, Stevenage, 2. ed. edition.

Wan, S. and Foxlin, E. (2010). Improved pedestrian nav-

igation based on drift-reduced mems imu chip. Pro-

ceedings of the 2010 International Technical Meeting

of The Institute of Navigation, pages 220–229.

Weinberg, H. (2002). Using the adxl202 in pedometer and

personal navigation applications. Technical report,

Analog Devices.

Zhao, N. (2010). Full-featured pedometer design realized

with 3-axis digital accelerometer. Analog Dialogue,

44(06).

SENSORNETS 2012 - International Conference on Sensor Networks

148