Using Motion-Awareness for the 3D Indoor Personal Navigation on a Smartphone

8
Using Motion-Awareness for the 3D Indoor Personal Navigation on a Smartphone Ling Pei, Ruizhi Chen, Jingbin Liu, Heidi Kuusniemi, Yuwei Chen, Tomi Tenhunen Finnish Geodetic Institute, Finland BIOGRAPHIES Dr. Ling Pei received his Ph.D degree in test measurement technology and instruments from the Southeast University, China, in 2007, joining the Finnish Geodetic Institute (FGI) at the same year. He is a Specialist Research Scientist in the Navigation and Positioning Department at FGI, where his research interests include indoor/outdoor seamless positioning, ubiquitous computing, wireless positioning, context awareness and location-based services. Dr. Ruizhi Chen is the Professor and Head of the Department of Navigation and Positioning at the Finnish Geodetic Institute. He holds a M.Sc. degree in computer science and a Ph.D degree in geodesy. His research interests include satellite-based augmentation systems, multi-sensor positioning, pedestrian navigation and mobile mapping systems. Dr. Jingbin Liu is a Specialist Research Scientist in the Department of Navigation and Positioning at the Finnish Geodetic Institute. Prior to joining FGI, he worked for more than four years as a GPS receiver firmware engineer at SiRF technology Inc. He received his M.Sc. and Ph.D degrees in geodesy in 2004 and 2008 from Wuhan University, China. His research interests cover various aspects of outdoor/indoor seamless navigation, including GNSS precise positioning, integrated GNSS/inertial sensor positioning, indoor location awareness based on wireless signals, software defined GNSS receiver technology, as well as GNSS-based meteorology. Dr. Heidi Kuusniemi is a Chief Research Scientist at the Department of Navigation and Positioning at the Finnish Geodetic Institute. She received her M.Sc. degree in 2002 and D.Sc. (Tech.) degree in 2005 from Tampere University of Technology, Finland. Her doctoral studies on personal satellite navigation were partly conducted at the Department of Geomatics Engineering at the University of Calgary, Canada. From 2005 to 2009 she worked as a GPS Software Engineer in research and development at Fastrax Ltd. Her research interests cover various aspects of GNSS and multi-sensor fusion algorithms for seamless outdoor/indoor positioning. Dr. Yuwei Chen received his B.Sc. from the Electronics Engineering Department of Zhejiang University (China 1999), M.Sc. from the Information and Electronic Department of Zhejiang University (China 2002), and a Ph.D in Circuit and System from the Shanghai Institute of Technical Physics (SITP), Chinese Academy of Science (China 2005). He is now working at the Finnish Geodetic Institute as a Specialist Research Scientist in the Department of Navigation and Positioning. Tomi Tenhunen has graduated as the Bachelor of Engineering (BEng) in Information Technology from HAMK University of Applied Sciences in 2008. His main subjects were computer networks and electronics. He is now working at the Department of Navigation and Positioning at the Finnish Geodetic Institute as an associate research scientist. His research interests include hardware and network designing, wireless positioning and location-based services. ABSTRACT The paper presents an indoor navigation solution combining physical motion recognition with wireless positioning in a three dimensional space. 27 features are extracted utilizing the built-in accelerometers and magnetometers in a smartphone. 8 common motion modes during indoor navigation, e.g., static, standing with hand swinging, normal walking with holding the phone in hand, normal walking with hand swinging, fast walking, U-turning, going up stairs, and going down stairs are detected by the three classification algorithms: Bayesian Network (BN), Decision Tree (DT), and Support Vector Machine (SVM) respectively. Test results indicate that the motion modes are recognized correctly up to 95.53% of test cases. A motion-awareness assisted wireless positioning approach is applied to determine the position of a smartphone user. The field tests show

Transcript of Using Motion-Awareness for the 3D Indoor Personal Navigation on a Smartphone

Using Motion-Awareness for the 3D Indoor

Personal Navigation on a Smartphone

Ling Pei, Ruizhi Chen, Jingbin Liu, Heidi Kuusniemi, Yuwei Chen, Tomi Tenhunen

Finnish Geodetic Institute, Finland

BIOGRAPHIES

Dr. Ling Pei received his Ph.D degree in test

measurement technology and instruments from the

Southeast University, China, in 2007, joining the

Finnish Geodetic Institute (FGI) at the same year. He

is a Specialist Research Scientist in the Navigation

and Positioning Department at FGI, where his

research interests include indoor/outdoor seamless

positioning, ubiquitous computing, wireless

positioning, context awareness and location-based

services.

Dr. Ruizhi Chen is the Professor and Head of the

Department of Navigation and Positioning at the

Finnish Geodetic Institute. He holds a M.Sc. degree

in computer science and a Ph.D degree in geodesy.

His research interests include satellite-based

augmentation systems, multi-sensor positioning,

pedestrian navigation and mobile mapping systems.

Dr. Jingbin Liu is a Specialist Research Scientist in

the Department of Navigation and Positioning at the

Finnish Geodetic Institute. Prior to joining FGI, he

worked for more than four years as a GPS receiver

firmware engineer at SiRF technology Inc. He

received his M.Sc. and Ph.D degrees in geodesy in

2004 and 2008 from Wuhan University, China. His

research interests cover various aspects of

outdoor/indoor seamless navigation, including GNSS

precise positioning, integrated GNSS/inertial sensor

positioning, indoor location awareness based on

wireless signals, software defined GNSS receiver

technology, as well as GNSS-based meteorology.

Dr. Heidi Kuusniemi is a Chief Research Scientist at

the Department of Navigation and Positioning at the

Finnish Geodetic Institute. She received her M.Sc.

degree in 2002 and D.Sc. (Tech.) degree in 2005

from Tampere University of Technology, Finland.

Her doctoral studies on personal satellite navigation

were partly conducted at the Department of

Geomatics Engineering at the University of Calgary,

Canada. From 2005 to 2009 she worked as a GPS

Software Engineer in research and development at

Fastrax Ltd. Her research interests cover various

aspects of GNSS and multi-sensor fusion algorithms

for seamless outdoor/indoor positioning.

Dr. Yuwei Chen received his B.Sc. from the

Electronics Engineering Department of Zhejiang

University (China 1999), M.Sc. from the Information

and Electronic Department of Zhejiang University

(China 2002), and a Ph.D in Circuit and System from

the Shanghai Institute of Technical Physics (SITP),

Chinese Academy of Science (China 2005). He is

now working at the Finnish Geodetic Institute as a

Specialist Research Scientist in the Department of

Navigation and Positioning.

Tomi Tenhunen has graduated as the Bachelor of

Engineering (BEng) in Information Technology from

HAMK University of Applied Sciences in 2008. His

main subjects were computer networks and

electronics. He is now working at the Department of

Navigation and Positioning at the Finnish Geodetic

Institute as an associate research scientist. His

research interests include hardware and network

designing, wireless positioning and location-based

services.

ABSTRACT

The paper presents an indoor navigation solution

combining physical motion recognition with wireless

positioning in a three dimensional space. 27 features

are extracted utilizing the built-in accelerometers and

magnetometers in a smartphone. 8 common motion

modes during indoor navigation, e.g., static, standing

with hand swinging, normal walking with holding the

phone in hand, normal walking with hand swinging,

fast walking, U-turning, going up stairs, and going

down stairs are detected by the three classification

algorithms: Bayesian Network (BN), Decision Tree

(DT), and Support Vector Machine (SVM)

respectively. Test results indicate that the motion

modes are recognized correctly up to 95.53% of test

cases. A motion-awareness assisted wireless

positioning approach is applied to determine the

position of a smartphone user. The field tests show

1.22 m mean error in the “Static Test” and 3.53 m in

the “Stop-Go Test”.

INTRODUCTION

Nowadays, with the explosive growth of capabilities

in handheld devices, various components are

embedded into mobile phones, such as GPS, WLAN

(a.k.a. Wi-Fi), Bluetooth, accelerometer,

magnetometer, camera, etc.

With the locating capability fast becoming one of the

standard features in mobile devices, more and more

people are getting used to the location-enabled life.

Employing the global navigation satellites system

(GNSS), the applications in the “smart” devices

greatly enrich the end users‟ outdoor activities.

However, given the nature of GNSS design, they are

clearly not primarily suited for the applications in

urban canyons and indoors. Satellite based

positioning technologies continue to struggle indoors

due to well known issues such as the weak signal or

non-line-of-sight (NLOS) conditions between the

mobile user and satellites.

Benefited from the existing infrastructure, the RF-

based technologies are definitely one of the most

potential alternatives. RADAR [3] is the first WLAN

based positioning system that computes the mobile

device‟s location based on radio signal strength

(RSS) from many access points (APs). Skyhook is a

system that depends on information about the AP‟s

coordinates in a database in order to predict location

[6]. Ekahau [7] provides an easy and cost effective

solution for locating people, assets, inventory and

other objects using Wi-Fi. The Active Badge [8]

system uses ceiling-mounted infrared sensor

detectors to detect signals from a mobile active

badge. Place Lab [9] has even more ambitious goals

by seeking to create a comprehensive location

database that uses fixed commodity Wi-Fi, GSM, and

Bluetooth devices as global beacons.

Meanwhile, the human physical activities recognition

by using MEMS sensors has been widely applied for

the health monitoring, emergency services, athlete

training, navigation, etc [10], [11]. Since the motion

sensors such as accelerometers and magnetometers

are integrated into the smartphone, it brings the

opportunity to assist navigation with the knowledge

about the motion of a pedestrian [12].

For the end users, the smartphone already contains

the potential for indoor navigation and positioning in

the existing infrastructures. This paper presents an

indoor pedestrian navigation solution relying on

motion-awareness in a three dimensional

environment where the WLAN infrastructure is

existing.

MOTIVES

Related researches indicate that utilizing signals of

opportunity e.g. WLAN, is an efficient locating

alternative in the GPS denied environment. However,

in order to minimize smartphone battery drain, the

WLAN scanning interval is always limited. For

instance, most of the Nokia mobile phones refresh the

scanned WLAN information around every 8-10

seconds. The default scanning interval of most

Android devices is 15 seconds.

During the positioning gaps where no wireless

signals are updated, the most important elements for

navigation are the movement speed and orientation.

As long as they are figured out, it is possible to

estimate the position of the user over time. Therefore,

this paper introduces the built-in accelerometer and

magnetometer on a smartphone to recognize the

user‟s movement information. The proposed solution

detects the physical movements by using simple

acceleration and orientation features throughout the

navigation process. With the recognized motions, it is

possible to estimate the reasonable position over the

period without wireless locating.

Most of the previous motion recognition related

researches assumed that the MEMS inertial sensors

are fixed on a human body [13],[14] (e.g. in a pocket,

clipped to a belt or on a lanyard) and an inference

model can be trained according to a handful of body

positions. Some of them use the phone as a sensor to

collect activities for an off-line analyzing purpose

[15]. Compared to the daily activities such as

„Sitting‟, „Walking‟,‟ Running‟, „Jumping‟, in three-

dimensional indoor structures, the motions of a

pedestrian using smartphone navigating are far more

complicated due to the arbitrary gestures while a

phone is kept in hand. Hence this literature primarily

focuses on the common motion modes of a user with

a phone in hand while navigating.

3D WIRELESS POSITIONING

The applicable wireless positioning methods include

Cell-ID, fingerprinting, trilateration, and utilizing an

Artificial Neural Network (ANN). Fingerprinting

[16],[17] as the most common approach is adopted in

this paper. The received signal strength indicators

(RSSIs) are the basic observables in this approach.

The whole process consists of a training phase and a

positioning phase. During the training phase, a radio

map of probability distribution of the received signal

strength is constructed for the targeted area. The

targeted area is divided into a matrix of grids, and the

central point of each grid is referred to as a reference

point. The probability distribution of the received

signal strength at each reference point is represented

by a Weibull function [4], and the parameters of the

Weibull function are estimated with the limited

number of training samples.

Based on the constructed radio map, the positioning

phase determines the current location using the

measured RSSI observations in real time. Given the

observation vector }...,{ 21 ksssS , the problem is

to find the most probable location (l) with the

maximized conditional probability )|( SlP ,

maximized by Bayesian theorem as

])(

)()|([maxarg)]|([maxarg

SP

lPlSPSlP ll (1)

Well-known node-arc model is employed to represent

a topological network of radio maps in three-

dimension. The topology dominates the connection

and distance between random two reference points.

The stairs are the only accesses to difference floors.

The grid based hidden Markov model (HMM) filter is

implemented to produce an optimal estimation based

on the previous state. The transit probability matrix

of HMM is computed according to the travelled

distance which can be estimated by the knowledge

about the motion modes over time [5]. For instance,

the travelled distance is zero if the current motion

mode is in static. The navigation user travels about

0.8 m per second while normally walking in an

indoor environment.

MOTION DEFINITION

Eight of the most common motion modes during

pedestrian navigating in a three-dimensional

infrastructure are considered in this paper. The

motion modes are grouped into four series as

following:

(1) S-series motion modes refer to the stationary

behavior during a navigating process. ST is a mode

where a user keeps a phone in hand without any

movement. In contrast, SS is a category of the

movements where the user‟s location does not change

but the phone in hand keeps swinging.

(2) W-series is relevant to walking. WH represents

the motion mode where the user is using a navigation

application on the handset while walking. WS stands

for the normal walking behavior where the user holds

the phone in hand with the hand swinging. WF is a

fast walking behavior with significant arm swinging.

(3) T-series is about turning. UT so called U-turning

is a spot turn without any displacement.

(4) V-series of the movements concern motions in

vertical dimension. US and DS are going up/down

the stairs.

FEATURE EXTRACTION

After the physical characteristics of the movement

behavior are analyzed, the collective features of

accelerometer (Feature 1-18, 22-27) and

magnetometer (Feature 19-21) in time domain

(Feature 1-21) and frequency domain (Feature 22-27)

are extracted for the motion mode estimation. Note

that, in Table 1, the dynamic acceleration denotes the

real-time acceleration reading from the smartphone

minus the gravity acceleration.

Table 1. Feature Definition

Feature

No.

Feature Name Feature Definition

1 MeanAccX Mean value of the

acceleration at x-axis.

2 MeanAccY Mean value of the

acceleration at y-axis.

3 MeanAccZ Mean value of the

acceleration at z-axis.

4 MeanAcc Mean value of the

acceleration.

5 MeanDynAccV Mean value of the

dynamic acceleration on

the vertical plane.

6 MeanDynAccH Mean value of the

dynamic acceleration on

the horizontal plane.

7 MeanAccH Mean value of the

horizontal acceleration.

8 MeanAccV Mean value of the

vertical acceleration

minus gravity

acceleration.

9 MeanDynAcc Mean value of the

dynamic acceleration.

10 VarAccX

Variance of the

acceleration at x-axis.

11 VarAccY Variance of the

acceleration at y-axis.

12 VarAccZ Variance of the

acceleration at z-axis.

13 VarAcc Variance of the

acceleration.

14 VarDynAccV Variance of the dynamic

acceleration on the

vertical plane.

15 VarDynAccH Variance of the dynamic

acceleration on the

horizontal plane.

16 VarAccH Variance of the

horizontal acceleration.

17 VarAccV Variance of the vertical

acceleration.

18 VarDynAcc Variance of the dynamic

acceleration.

19 MeanMag Mean value of the

heading.

20 DiffMag Heading change.

21 VarMag Variance of the heading.

22 1stFreqAcc 1st dominant frequency

of the acceleration.

23 Amp1stFreqAcc Amplitude of the1st

dominant frequency of

the acceleration.

24 2ndFreqAcc 2nd dominant frequency

of the acceleration.

25 Amp2ndFreqAcc Amplitude of the 2nd

dominant frequency of

the acceleration.

26 FreqDiffAcc Difference between two

dominant frequencies.

27 AmpScaleAcc Amplitude scale of two

dominant frequencies.

MOTION RECOGNITION

The motion recognition aims at determining which of

the eight motion modes have effectively caused the

above 27 features. The classification algorithms such

as k-Nearest Neighbour (kNN), Linear Discriminant

Analysis (LDA), Quadratic Discriminant Analysis

(QDA), Naïve Bayesian Classifier (NBC), Bayesian

Network (BN), Decision Tree (DT), Artificial Neural

Networks (ANN), Support Vector Machine (SVM)

and so forth can be applied. In this paper we

implemented BN, DT, and SVM for comparison.

Bayesian Network

Bayesian network is a graphical model that encodes

probabilistic relationships among variables of

interest. A Bayesian network can be used to learn the

causal relationships, and hence can be used to obtain

the understanding about a problem domain and to

predict the consequences of intervention. Bayesian

network models encode the strength of causal

relationships with probabilities. Consequently, prior

knowledge and data can be combined with well-

studied techniques from Bayesian statistics.

The Bayesian network can be represented as various

models such as Gaussian mixture model (GMM),

Principal Components Analysis (PCA), Hierarchical

mixtures of experts (HME), Quick Medical

Reference (QMR), Conditional Gaussian model,

other hybrid models, etc.

In this paper, we implement GMM in which eight

modes of motion and 27 features are represented by a

multi-modal mixture of unimodal Gaussians. The

model is trained by using the Expectation

Maximization (EM) algorithm. EM works by starting

with a randomly initialized model, and then it

iteratively refines the model parameters to produce a

locally optimal maximum-likelihood fit. The EM

algorithm is composed of two steps. In the first, each

data point undergoes a soft-assignment to each

mixture component. In the second, the parameters of

the model are adjusted to fit the data based on the soft

assignment of the previous step.

Decision Tree

The collected data indicate that the relationship

between labelled motions and extracted features is

not suitable to approximate by a linear model.

Furthermore, it is hard to figure out a form for the

nonlinear least squares regression. Decision tree as a

nonparametric type of regression fitting approach is

adopted for predicting the motion given a set of fitted

response values and observables. The tree structure

and the fitted response values are trained from the

training datasets.

Given a set of the feature measurements and a setup

tree, the algorithm asks whether the measurements

satisfy a given condition. Depending on the answers

to one question, the algorithm either proceeds to

another question or arrives at a fitted response value

for a particular motion mode.

Support Vector Machine (SVM) Generally, a SVM is a so-called maximum margin

classifier. The objective of the SVM optimization

problem is to obtain certain parameters in order to

define a separating hyperplane that has optimal class

separability (optimal in terms of maximum margin

that is defined by the support vectors). In real world

scenarios it often happens that features are close to

the hyperplane or cannot be separated properly.

Therefore slack variables are introduced (soft side

constraints). These allow for a certain amount of

misclassified features.

If the data is not linearly separable, the so-called

kernel trick comes into play. There are different kinds

of kernels such as polynomial kernel, (Gaussian)

radial basis, and sigmoid functions. Kernels

transform the original data into a higher dimensional

feature space. Even if the original data are nonlinear,

the transformed data is separable by a hyperplane in

feature space.

PEDESTRIAN DEAD RECKONING

ALGORITHM

From the training data, the velocity of each motion is

estimated as shown in Table 2. The speed of ST, SS,

and UT is zero because the user does not move in

these states. The velocities for WH, WS, WF, US,

and DS were trained at a corridor and stairs with

known length. Testers made marks of time stamps at

the beginning and the end of each motion training

data set. Velocity is calculated by dividing the known

length of walking by its duration.

Table 2. Velocity estimated

Motion

modes

Velocity (m/s)

ST 0

SS 0

WH 0.76

WS 0.80

WF 1.80

UT

US

DS

0

0.55

0.6

User‟s displacement each second can be calculated as

tVS motion (2)

where motionV is the velocity output based on the

recognized motion mode and t is the time.

The pedestrian dead reckoning (PDR) solution

utilizes the displacement of the user, and is

implemented in a Local-Level Frame (LLF). The

positioning equations are rather straightforward as

follows [18]:

)sin(

)cos(

1

1

kkk

kkk

SXX

SYY

(3)

where k denotes the current epoch, Y is the coordinate in east direction, X is the coordinate in north direction, S is the distance travelled by the user, and φ is the heading.

The PDR positioning algorithm includes the following procedures:

Motion detection,

Speed estimation,

Travelled distance calculation,

Determination of heading, and

Positioning.

In the experiments described in the following section,

the heading is obtained from a magnetometer.

However, the magnetometer has some important

drawbacks. Indeed, magnetic disturbances are

numerous, particularly in indoor environments.

Hence we applied magnetometer readings after map

matching instead of using the heading directly. With

the heading input from the magnetometer and current

position estimate, matched direction is derived from

the segment vector in the topological network of the

fingerprint database. In addition, the cumulated

travelled distance over the WLAN positioning

absence duration will be used as an observation in the

HMM grid filter for determining the position.

EXPERIMENT RESULTS

To verify the solution proposed in this paper, some

field tests were carried out in the Finnish Geodetic

Institute (FGI) office building with three floors. The

8 motion modes described earlier were involved in

the tests. A tester performed 8 motions inside the FGI

building over 40 minutes. A smartphone based

application was developed for collecting sensor data,

labelling the motion mode, and locating the

smartphone position. The collected data were

grouped into two datasets: training set and testing set,

where all the features are extracted over a 3 seconds

window. The training set was used for learning the

parameters of three classification algorithms: BN

(GMM), DT, and SVM. The testing set with labelled

motion was used for motion recognition by applying

each classification algorithm. The results indicate that

a 95.53% recognition rate is possible with the use of

the SVM classifier, 83.80% with the DT classifier,

and only 73.74% with the BN classifier.

Table 3. Classifier vs. Accuracy

Classifier SVM BN DT

Accuracy 95.53% 73.74% 83.80%

Figures 1-3 illustrate the motion predictions by

applying BN (GMM), DT, and SVM classifiers,

respectively. Figure 1 indicates that the BN GMM

classifier can predict the S-series and W-series

motion modes with quite high accuracy. In contrast,

T-series and V-series motion modes are merely

recognized. Figure 2 shows that the recognition rates

of T-series and V-series motion modes are improved

by using the DT classifier. However, it is evident that

much more confusion occurred between ST and UT.

In addition, the low ability to discriminate US from

DS suggests that it might be possible to improve the

recognition rate by grouping “Using-the-stairs”

motions. This will reduce the effect of confusion

between DS and US. Then crossing the stairs can be

used for floor detection. Figure 3 presents the

significant improvement of motion recognition rate

when the SVM classifier is adopted. The confusion

matrix for motion recognition is listed in Table 4.

Although 18.75% of ST motions and 22.22% of US

motions are taken for UT, the SVM classifier has an

efficient performance of motion recognition rate

compared to the BN and DT classifiers.

0 20 40 60 80 100 120 140 160 1801

2

3

4

5

6

7

8Baysian Network Classifier

Samples

Motion

Labelled Motion

Estimated Motion

Figure 1. BN (GMM) Motion Predictions (Motion

Mode 1:ST, 2:SS, 3:WH, 4:WS, 5:WF, 6:UT, 7: US,

8:DS)

0 20 40 60 80 100 120 140 160 1801

2

3

4

5

6

7

8Decision Tree Classifier

Samples

Motion

Labelled Motion

Estimated Motion

Figure 2. DT Motion Predictions (Motion Mode

1:ST, 2:SS, 3:WH, 4:WS, 5:WF, 6:UT, 7: US, 8:DS)

0 20 40 60 80 100 120 140 160 1801

2

3

4

5

6

7

8SVM Classifier

Samples

Motion

Labelled Motion

Estimated Motion

Figure 3. (LS-)SVM Motion Predictions (Motion

Mode 1: ST, 2: SS, 3:WH, 4:WS, 5:WF, 6:UT, 7: US,

8:DS)

Table 4. Confusion Matrix for the motion recognition

from SVM classifier (Unit: %)

Recognized Motion

ST SS WH WS WF UT US DS

La

be

lle

d M

oti

on

ST 81.25 0 0 0 0 18.75 0 0

SS 0 100 0 0 0 0 0 0

WH 0 0 100 0 0 0 0 0

WS 0 0 0 100 0 0 0 0

WF 0 0 0 0 100 0 0 0

UT 0 0 0 0 0 100 0 0

US 0 0 0 0 0 22.22 77.78 0

DS 0 0 0 0 0 0 0 100

To prove the advantage of wireless positioning by

using motion-awareness, two positioning tests were

conducted in the FGI building. The first test called

“Static Test” was carried out in the static state. A user

stood on the reference point while holding the phone

in hand (ST) for ten minutes. The results are

summarized in Table 5 where the average error is

3.43 m when the Maximum Likelihood (ML)

algorithm was applied for the wireless positioning.

The mean error is reduced to 1.22 m if applying a

motion-awareness assisted HMM algorithm.

Furthermore, compared to the ML algorithm, the

RMSE (Root Mean Square Error) and maximum

error are all significantly decreased when using the

motion-awareness assisted HMM algorithm.

Table 5. Static Test (Unit: m)

Static Test ML HMM (Motion-

Awareness)

Mean 3.43 1.22

RMSE 5.98 2.55

MaxErr 21 9

MinErr 0 0

The second test is called “Stop-Go Test”. In the FGI

office building, a tester stopped at each reference

point to obtain the wireless positioning estimation,

then moved to another reference point while

randomly performing motions between two stops.

Table 6 compares the positioning results derived

from ML and motion-awareness assisted HMM

algorithm. Improvements can be found in motion-

awareness assisted HMM algorithm compared to the

ML algorithm. Meanwhile, the motion awareness

also raises the floor detection rate from 89.93% to

95.95%. The details are shown in Tables 7 and 8.

Table 6. Stop-Go Test (Unit: m)

Stop-Go Test ML HMM (Motion-

Awareness)

Mean 4.38 3.53

RMSE 6.02 4.55

MaxErr 18 9

MinErr 0 0

Table 7. Confusion matrix for floor detection by

using ML wireless positioning

Estimated

Floor Labelled Floor

1st 2nd 3rd

1st 93.94 % 6.06 % 0 %

2nd 4.00 % 92.00 % 4.00 %

3rd 0 % 17.95 % 82.05 %

Table 8. Confusion matrix for floor detection by

using motion-awareness HMM wireless positioning

Estimated

Floor Labelled Floor

1st 2nd 3rd

1st 96.97 % 3.03 % 0 %

2nd 4.00 % 96.00 % 0 %

3rd 0 % 5.13 % 94.87 %

CONCLUSION AND FUTURE WORK

In this paper, a motion-awareness assisted wireless

positioning method is presented. The raw data from

accelerometer and magnetometer on a smartphone are

converted into 27 features. The eight motion modes

are predicted by applying BN (GMM), DT, and (LS-)

SVM classifiers, respectively. The test results

indicate that the SVM classifier has an efficient

performance of motion recognition rate compared to

BN and DT classifiers. The recognition rates of T-

series and V-series motions are lower than those of S-

series and W-series motions. Furthermore, both

positioning accuracy and floor detection rate are

significantly improved by applying motion-awareness

in the wireless positioning algorithms.

The motion mode recognition solution proposed in

this paper provides motion recognition rate up to

95.53% of the test cases. However, the motion

behavior varies from person to person. In the future

we will involve more persons for testing the motion

recognition algorithms and determine the most useful

features for classification. Besides, more motions will

be considered for indoor and outdoor navigation. For

instance, we only consider the “using-stairs” motion

in the V-series motions. The other V-series motions

such as “using-elevator” will also be researched. In

addition, the T-series motions introduce much more

confusion because they are possible to be conducted

with other motions simultaneously. Therefore, more

effort will be concentrated on the T-series motions.

Additionally, in this paper, the velocity of each type

of motion is trained offline. This introduces speed

estimation errors when the user changes his/her

walking behaviour. In future research, we will make

full use of the acceleration features to estimate the

speed online.

ACKNOWLEDGMENTS

This work was a part of the project INOSENSE

(INdoor Outdoor Seamless Navigation for SEnsing

Human Behavior) funded by the Academy of Finland

with the Finnish Geodetic Institute, Department of

Navigation and Positioning, and University of

Helsinki, Department of Psychology.

REFERENCES

[1] I. Kraemer and B. Eissfeller. A-GNSS: A

Different Approach.InsideGNSS. Vol 4 (5): pp. 52-61. September 2009.

[2] Canalys Q3 2008 Research, 2008. Avialable at http://www.canalys.com/pr/2008/r2008111.html. visited on 4 September 2010.

[3] P. Bahl and V. Padmanabhan, Radar: A In-Building RF Based User Location and Tracking System, Proc. IEEE INFOCOM, pp. 775-784, Mar. 2000.

[4] Ling Pei, Ruizhi Chen, Jingbin Liu, Heidi Kuusniemi, Tomi Tenhunen, Yuwei Chen, Using Inquiry-based Bluetooth RSSI Probability Distributions for Indoor Positioning, Journal of Global Positioning Systems (2010), Vol.9, No.2 :122-130.

[5] J. Liu, R. Chen, L. Pei, W. Chen, T. Tenhunen, H. Kuusniem, T. Kröger, Y Chen. (2010), Accelerometer Assisted Wireless Signals Robust Positioning Based on Hidden Markov Model, Proceedings of the IEEE/ION position, location and navigation symposium (PLANS) 2010, Indian Wells, CA, USA, pp. 488 – 497.

[6] K. Pahlavan, F. Akgul , Y. Ye, T. Morgan, F. Alizadeh-Shabdiz, M. Heidari and C. Steger. Taking Positioning Indoors Wi-Fi Localization and GNSS. InsideGNSS. Vol 5 (3): pp. 40-47. May 2010.

[7] Ekahau.Inc., http://www.ekahau.com/, visited on 4 September 2010.

[8] Roy Want, Andy Hopper, Veronica Falcão, Jonathan Gibbons, The active badge location system, ACM Transactions on Information Systems (TOIS), v.10 n.1, pp.91-102, Jan. 1992.

[9] B.Schilit, A.LaMarca, G.Borriello, William Griswold D. M., Lazowska E., Balachandran A., and Iverson J. H. V (2003). Challenge: Ubiquitous location-aware computing and the Place Lab initiative. In First ACM International

Workshop of Wireless Mobile Applications and Services on WLAN, September 2003.

[10] Frank K., M. J. V. Nadales, P. Robertson, and M. Angermann (2010) “Reliable real-time recognition of motion related human activities using MEMS inertial sensors”, in Proc. of the 23rd International Technical Meeting of The Satellite Division of the Institute of Navigation, Portland, OR, September

[11] Susi, M., D. Borio and G. Lachapelle (2011) Accelerometer Signal Features and Classification Algorithms for Positioning Applications. Proceedings of International Technical Meeting, Institute of Navigation, San Diego, 24-26 January.

[12] Pei L., R. Chen, J. Liu, W. Chen, H. Kuusniemi, T. Tenhunen, T. Kröger, Y. Chen, H. Leppäkoski and J. Takala (2010) “Motion recognition assisted indoor wireless navigation on a mobile phone”, in Proc. of the 23rd International Technical Meeting of The Satellite Division of the Institute of Navigation, Portland, OR, September, pp. 3366–3375 2010.

[13] Kavanagh JJ, Menz HB. Accelerometry: A technique for quantifying movement patterns during walking. Gait Posture, 28 (1), pp. 1-15, July 2008.

[14] Baek, J., et al. (2004): Accelerometer signal processing for user activity detection, Knowledge-Based Intelligent Information and Engineering Systems, LNAI 3215, pp. 610-617.

[15] J. Yang, Toward Physical Activity Diary: Motion Recognition Using Simple Acceleration Features with Mobile Phones, The 1st International Workshop on Interactive Multimedia for Consumer Electronics (IMCE), ACM Multimedia 2009. pp. 1-10.

[16] M. Youssef, A. Agrawala, and A. U. Shankar. WLAN location determination via clustering and probability distributions. In Proc. the First IEEE International Conference on Pervasive Computing and Communications (PerCom 2003), pp.143-150. IEEE Computer Society, Texas, USA, March 2003.

[17] Roos T., Myllymaki P., Tirri H., Misikangas P., and Sievänen J. (2002), A probabilistic approach to WLAN user location estimation, International Journal of Wireless Information Networks, 9(3),pp.155-164, July 2002.

[18] W. Chen, Z. Fu, R. Chen, Y. Chen, O. Andrei, T. Kröger, and J. Wang, “An integrated GPS and multi-sensor pedestrian positioning system for 3D urban navigation,” Proc. Urban Remote Sensing Event, 2009 Joint, Shanghai, China.