TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, … · Human Mobility Enhances Global Positioning Accuracy for Mobile Phone Localization Chenshu Wu, Student Member, IEEE, Zheng

TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. X, NO. X, SEPTEMBER 2013 1

Human Mobility Enhances Global PositioningAccuracy for Mobile Phone Localization

Chenshu Wu, Student Member, IEEE, Zheng Yang, Member, IEEE, Yu Xu, Member, IEEE, YiyangZhao, Member, IEEE, and Yunhao Liu, Senior Member, IEEE

Abstract—Global Positioning System (GPS) has enabled a number of geographical applications over many years. Quite a lot oflocation-based services, however, still suffer from considerable positioning errors of GPS (usually 1m to 20m in practice). In thisstudy, we design and implement a high-accuracy global positioning solution based on GPS and human mobility captured by mobilephones. Our key observation is that smartphone-enabled dead reckoning supports accurate but local coordinates of users’ trajectories,while GPS provides global but inconsistent coordinates. Considering them simultaneously, we devise techniques to refine the globalpositioning results by fitting the global positions to the structure of locally measured ones, so the refined positioning results are morelikely to elicit the ground truth. We develop a prototype system, named GloCal, and conduct comprehensive experiments in bothcrowded urban and spacious suburban areas. The evaluation results show that GloCal can achieve 30% improvement on averageerror with respect to GPS. GloCal uses merely mobile phones and requires no infrastructure or additional reference information. As aneffective and light-weight augmentation to global positioning, GloCal holds promise in real-world feasibility.

Index Terms—GPS, mobile phone localization, human mobility

F

1 INTRODUCTION

G Lobal positioning technology has enabled a greatnumber of yet-unimagined applications and attract-ed millions of civil users worldwide. Among all posi-tioning techniques, Global Positioning System (GPS) [1]is widely adopted from industries such as aviation, nau-tical navigation, and land surveying, to personal appli-cations such as driving navigation, object and individualtracking, and location sharing. Along with the popularityof mobile phones with built-in GPS, location informa-tion is available in more people’s pockets. Burgeoningmarkets of mobile phone applications such as location-based social networks, geocaching, and geotagging, etc.,are telling a true success story of the integration of GPSand mobile phones.

Although GPS has proven its availability and depend-ability over many years, many location-based servicesstill suffer from considerable errors of GPS. Albeit theofficially reported accuracy with high-quality GPS re-ceivers can achieve 3 meters [2], the actual accuracy usersattain from commodity smartphones ranges from 1m toup to 20m, which limits the uses of numerous applica-tions, leaving room for various augmented technologies.

Generally, GPS accuracy is affected by a number ofunavoidable factors, including satellite positions, atmo-spheric conditions, and the blockage to the satellitesignals caused by mountains and buildings, etc. Toovercome or bypass these factors, several augmentationsystems, for instance, Assisted GPS (AGPS) [3], Differ-ential GPS (DGPS) [4], and Wide Area Augmentation

Chenshu Wu, Zheng Yang, Yiyang Zhao, and Yunhao Liu are with the Schoolof Software and TNList, Tsinghua University. Yu Xu is with the WenzhouUniversity. E-mail: {wu, yang, yxu.wzu, zhaoyy, yunhao}@greenorbs.com

System (WAAS) [5], have been developed to aid GPS byproviding accuracy, integrity, availability, or any otherimprovement that is not inherently part of GPS itself.

Conventional augmentation systems mostly rely onfixed reference locations, e.g., cell towers, and hencerequire specific infrastructure provided by either publicor private sectors. Consequently, it is difficult for mobilephone users to embrace these augmentations any time atany place. Motivated by the proliferation of smartphoneswith rich internal sensors, we propose to enhance the ac-curacy of global positioning technology by utilizing localposition information captured by only mobile phones.

Nowadays, mobile phones possess powerful compu-tation and communication capability, and are equippedwith various functional built-in sensors. These sensorsenable so-called inertial sensing to characterize humanmobility [6], [7]. Inertial sensing, a.k.a. dead reckoning,is means of calculating one’s current location by usinga previously determined location and the estimations ofdisplacement and direction moved. With internal sensorslike accelerometer, gyroscope, and compass (or magne-tometer), which, respectively, reveal the acceleration, ro-tational velocity, and direction of user motion, one user’smoving trajectory can be tracked by dead reckoning [8].

Phone-based dead reckoning supports accurate butlocal coordinates of users’ trajectories, while GPS pro-vides global but inconsistent coordinates. This studyaims at bridging phone-based dead reckoning and GPSto offer high-accuracy global positioning. We presentGloCal (naming thanks to its connotation of ‘thinkGLObally and act loCALly’), a global positioning refine-ment approach via local trajectories tracked by mobilephones. The rationale behind GloCal is that global po-sitions can be refined by fitting their structure to that

2 TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. X, NO. X, SEPTEMBER 2013

of local positions, which is more accurate and henceeliciting the ground truth (Fig. 1a). To faithfully depictusers’ trajectory, GloCal first employs a novel scheme forhighly accurate dead reckoning on mobile phones. Thedead-reckoned local trajectory and the global trajectory,obtained from a series of GPS measurements, are thenconverted to local and global coordinates in a 2D plane,respectively. On this basis, GloCal refines the originallyinaccurate global positions by transforming the localcoordinates into the global ones using a set of translation,scaling, and rotation operations, as illustrated in (Fig. 1b).

To evaluate our design, we implement a prototype onAndroid OS using Google Nexus S phones and conductcomprehensive experiments in both crowded urban andspacious suburban areas. The evaluation results suggestthat GloCal can reduce 30% of global positioning errorsof GPS with only negligible extra energy consumption,which demonstrates the feasibility of GloCal in realworld deployment.

The major contributions are as follows.

• We propose a novel approach to improve glob-al positioning accuracy using local trajectories de-lineated by mere mobile phones. GloCal requiresneither additional infrastructure nor fixed referencepoints. GloCal works with one user using his mobilephone while walking in normal course, exerting noadditional constraints to users. Besides GPS, othercoarse-grained global positioning techniques likeGSM- and WiFi-based localization can also benefitfrom this design.

• We introduce a new scheme for smartphone-enableddead reckoning, which achieves precise step count-ing, stride estimation, and direction reckoning, with-out any dependence on extra information such asdigital maps or floor plans.

• A coordinate transformation algorithm is intro-duced to achieve effective transformation betweenlocal and global coordinate system, which is agnos-tic to the specific localization techniques used. Inother words, given two groups of localization resultswith different accuracy, the algorithm is universallycapable of improving the precision of the less accu-rate one in its coordinate system.

• We implement a prototype on commodity mobilephones and conduct real world evaluation in bothurban and suburban areas. The results show thatGloCal greatly reduces the average error of GPSfrom 5∼10m to around 3m, which was indicatedempirically possible but only with dedicated infras-tructure or high-quality receivers.

The rest of this paper is organized as follows. Section 2introduces the system design of GloCal. A novel schemefor dead reckoning, as well as the local and globalcoordinate generation, is presented in Section 3. Sec-tion 4 illustrates how to transform the local coordinatesystem to the global one. In Section 5, we provide theexperiments and evaluations. We review related works

G

G

G

G

GG

G

G

G

GG

G

G

G G

G

G GPS

Footprints

Ground Truth

(a) An illustration of user trajectory with both GPS and local measure-ments. For the ease of visualization, distances between footprints arelarger than the facts.

G Global Positions

Local Positions

G

G G

G

G

G

G

G

G

G

G

GG

G G

G

(b) Refine global positions with transformed local coordinates

Fig. 1. Global positions refined by local footprints elicitthe ground truth fantastically well.

in Section 6 and conclude the work in Section 7.

2 OVERVIEW AND CHALLENGES

We first present the system architecture of GloCal, fol-lowed by the challenges faced.

As shown in Fig. 2, the working process of GloCal con-sists of two core phases: coordinate generation and co-ordinate transformation. Imagine that when a user usesnavigation in some scenic, he may turn to global posi-tioning services for desired locations using his mobilephone. At this point, apart from the global locations,GloCal records the internal sensor readings of the mobilephone that is equipped with accelerometer, gyroscope,and compass, etc. These consecutive global locationsalong the user trajectory form a global coordinate system,while the local measurements from inertial sensors willconstruct a local coordinate system.

GloCal characterizes and exploits user mobility toattain local position information. Analogous to con-ventional dead reckoning techniques, GloCal leveragesvarious sensors to infer user walking characteristics andfurther to depict the entire user trajectory. Specifically,GloCal uses accelerometer to identify user walking stepsand gyroscope to estimate moving directions. Acceler-ation feature is further investigated to determine theaccurate stride length of a specific user. The walking dis-placement is then derived by multiplying the step countswith the stride length. Provided that the displacementand direction are available, a user trajectory beyond GPS

WU et al.: HUMAN MOBILITY ENHANCES GLOBAL POSITIONING ACCURACY FOR MOBILE PHONE LOCALIZATION 3

Coordinate Transformation

Local Coordinate Generation

Global Coordinate Generation

Displacement Ranging

Direction Reckoning

Accelerometer Raw Global Positions

Enhanced Global Positions

Gyroscope

Fig. 2. System architecture of GloCal

is obtained and a local coordinate system, namely, therelative locations, is accordingly delineated.

Observing that the dead-reckoned local positions pre-serve the structure of ground truth trajectory better thanthe global positions obtained from GPS, GloCal thereforeintends to improve the global positioning accuracy byfitting the global positions to the local ones. The bestfitting is achieved by realizing an optimal transformation,including a set of translation, scaling, and rotation opera-tions, that converts the local coordinates exactly into theglobal ones, minimizing the sum of squares of residualerrors. All global positions along the user trajectory areconcurrently refined with the transformed local positionsonce the optimal transformation is accomplished.

The intuitive idea of GloCal involves great challenges:

1) Given the fact that the internal sensors are noisyand that even tiny errors might be rapidly magni-fied by integration, realizing accurate dead reckon-ing is non-trivial. While direct integration over timeis subject to accumulative errors and is proved tobe unfeasible in practice, the method of step countspredominant in the literature is an alternative [6],[9]. However, the difficulty still remains since thatthe stride lengths vary from user to user and fromscenario to scenario.

2) Since the two coordinate systems are constructedby measurements from different techniques andthus are completely independent from each other,how to benefit from the local position informationto upgrade the global positioning accuracy?

The following section details our novel dead reckoningscheme and Section 4 addresses the challenges in har-nessing the local coordinates.

3 COORDINATE GENERATION

In this section, we first present a novel scheme forsmartphone-enabled dead reckoning to depict users’traveling trajectory. On this basis, the generation of localand global coordinates is introduced, respectively.

S1S2 S3

S6 S5

a>neg

a≤nega≤negPeak

a≥posPeak

a≤nega>negPeak

a≥pos

a>neg

S8

a≥posPeak

a>pos & a


0 1 2 3 4 5 6

−6

−4

−2

0

2

4

6

8

10

Time (s)

Acce

lera

tio

n (

m/s

2)

Raw data

Filtered data

Step starting

Step ending

Fig. 4. Results of FSM based stepcounting algorithm

−5

0

5

10mean=−0.08, std=1.08 mean=−0.06, std=1.11stride=60cm

−5

0

5


0 1 2 3 4 5 6 7−5

0

5


Time (s)

Acc

eler

atio

n am

plitu

de (

m/s

2 )

Fig. 5. Walking patterns of users withdifferent strides

O A B

djA (xj-1,yj-1)

B (xj,yj)

x

y

Fig. 6. Local coordinate system gen-eration

with different stride lengths exhibit considerable differ-ent characteristics such as variance, yet evince similarlyrepetitive patterns. We validate this design on real usersin Section 5 and the results indicate reasonable accuracy.

3.1.2 Direction ReckoningA lot of previous work leverage compass to estimateuser orientation, while some also take gyroscope intoconsideration. Compass reveals the absolute orientationrelative to the surface of the earth conveniently, butis usually pretty noisy. [8], [10]. Since GloCal solelyleverages user mobility to construct a relative coordinatesystem, the absolute orientation is not necessaryinvolved. Consequently, GloCal is free of using thenoisy compass and employs solely the gyroscope, whichprovides accurate angular velocity of human motion,to infer the changes of direction during every step.The method is fairly intuitive: integrating the angularvelocity captured by the gyroscope with respect to timewithin the interval of a step (detected by the FSM-basedalgorithm). Due to the high sampling frequency ofgyroscope (about 800Hz with our experimental phones)and its insensitivity to magnetic fields, the directionchanges can be precisely estimated and hence thestructure of a user path can be well identified, which isexactly what GloCal desires.

Dead-reckoning is well known to suffer from accumu-lative errors over time [7], [10], [12]. We have significant-ly reduced such accumulative errors by employing stepcounting and stride length estimation, which are bothimmune to cumulative errors. The direction estimation,however, still experiences accumulative errors over time.Although the accumulative errors in direction estimationcan be calibrated by leveraging digital compass [8], weshy away them by using only short trajectories thatare free from severe accumulative errors for coordinatetransformation (detailed in next section). For excessivelylong traces, we break down them into shorter parts andemploy piecewise operations.

3.2 Local Coordinate SystemIf the displacement and changes of direction of eachstep are known, we can build a Cartesian coordinatesystem, namely, the local coordinate system, to portray

the trajectory. Given a trajectory S = {s1, s2, · · · , sN}of N steps, each step sj corresponds a displacementdj and a direction change γj . Treating each step as apoint and the start of the first step s1 as the origin withcoordinates (0, 0), the coordinate of each point can beobtained, where the direction of the vector from s1 to s2is defined as that of the x axis and the orthogonal vectoris y axis. As shown in Fig. 6, assuming the coordinates ofstep sj−1 (point A) is (xj−1, yj−1), then the coordinatesof the next step sj (point B) can be calculated as

(xj , yj) = (xj−1+dj cos(φ+γj), yj−1+dj sin(φ+γj)), (1)

where φ =∑j−1

p=1 γp is the separation angle of vector ~OAand the x axis. Noting that in GloCal the γj is negativeif the direction change is clockwise, otherwise positive.

Actually, it is not necessary to calculate the coordinatesof each step in practice. Since the energy-hungry GPSis usually not such frequent as step rate, we only needto take into account those steps which are accompaniedwith global positioning stamps, which results in lowercomputation complexity and fewer energy cost.

3.3 Global Coordinate SystemGlobal positioning technology reports geographical lo-cations on the spherical surface of the earth. Such geo-graphical coordinates are usually in the form of (λ, ψ)(without loss of rigor, we consider the elevation to bezero), where λ and ψ denote the longitude and latitude(in degrees), respectively. As we assume the global co-ordinate system and the local one are co-planar, suchgeographical coordinates must be converted into 2DCartesian coordinates.

Fortunately, GPS reported geographic coordinates canbe accurately converted to Universal Transverse Merca-tor Grid System (UTM) format [13]. UTM is a formal,globally referenced planimetric coordinate system, and isalso the most common map standard today (supportedby most GPS receivers). In UTM, a point is located byspecifying a hemispheric indicator, a zone number, aneasting value, and a northing value. The coordinates arein the form of (E,N), where E and N denotes the eastingand northing values (in meters), respectively. In GloCal,we convert all GPS readings in the form of longitudeand latitude to the UTM format based on formulasmentioned in [13] for further processing.


4 COORDINATE TRANSFORMATIONAt this point, we have obtained both the local andglobal coordinates. In the following, we present how toimprove the global positioning accuracy by harnessinglocal positions. Our method is based on transforming thelocal coordinate system into the global one using a setof translation, scaling, and rotation operations based onHorn’s method [14].

Horn presented a closed-form solution of absolute ori-entation problem using unit quaternions in 2D and 3Dspace in [14]. Absolute orientation [15] is referred tofinding the relationship, i.e., recovering the transforma-tion, between two coordinate systems using pairs ofthe coordinates of a number of points in both systems,which is a classical problem in photogrammetric andin robotics. Horn’s solution uses unit quaternions torepresent rotation in 3D space. In GloCal, assumingthe local and global coordinate systems are both in aplane, unit quaternion is not necessary used. Instead,we use complex numbers to denote the coordinates ofpoints, for which the rotation can be represented as amultiplication between numbers, and derive a form ofoptimal transformation.

4.1 Problem FormulationSince the local trajectory will generally contain muchmore points than the GPS sampling points, (i.e., theglobal coordinate counts), and the two coordinate sys-tems are obviously independent from each other beforethe transformation is done, it is necessary to align thepoint number as well as to associate the correspondingpoints between the two coordinate set. In GloCal, thepoint number is preferentially determined by the globalcoordinate system. The local coordinates are then alignedby filtering the timestamps. That is, for each globalpoint, the local point which has the closest timestampis selected as the associated point. By doing this, bothcoordinate systems consist of the same number of one-to-one corresponding points.

Assume there are n points in the local coordinatesystem, denoted as L = {wj , j = 1, . . . , n}, and n corre-sponding points in the global coordinate system, denot-ed as G = {zj , j = 1, . . . , n}. Instead of a 2-dimensionalvector, each point is represented as a complex number,i.e., zj = zx,j + izy,j , wj = wx,j + iwy,j . According to[14], the transformation between these two coordinatesystems L and G can be thought of a rigid-body motionand can thus be decomposed into a translation, a scaling,and a rotation. In other words, the problem is to look fora transformation of the form

wg = sR(wl) + t0 (2)

from the local to the global coordinate system, wherewl ∈ L, wg is the corresponding transformed one inglobal coordinate system, s is a scale factor, t0 is thetranslational offset, and R(wl) denoted the rotated ver-sion of wl. Unless the data are perfect, we will not be able

to find a transformation such that the equation above issatisfied for each pair of points in L and G. Hence, theoptimal solution aims to minimize the sum of squaresof the residual errors:

n∑j=1

‖ ej ‖2=n∑

j=1

‖ zgj −wgj ‖

2, (3)

where zgj ∈ G and ej is the residual error between zgj

and wgj .As Horn’s solution does, we consider the total residual

errors first with translation, then with scaling, and finallywith respect to rotation.

4.2 TranslationFirst of all, we refer all positions to centroids defined by

z̄g =1

n

n∑j=1

zgj , w̄l =

1

n

n∑j=1

wlj , (4)

and derive the following new coordinates: zḡj = zgj −

z̄g, wl̄j = wlj−w̄l.Note that

∑nj=1 z

ḡj = 0,

∑nj=1 w

l̄j =

0.The residual error can be rewritten as

ej = zḡj −w

ḡj = z

ḡj − sR(w

l̄j)− t̄0, (5)

where t̄0 = t0 − z̄g + sR(w̄l). The sum of squares of theresiduals becomes

n∑j=1

‖ zḡj − sR(wl̄j) + t̄0 ‖2

=

n∑j=1

‖ S ‖2 +2t̄0 ·n∑

j=1

S + n ‖ t̄0 ‖2(6)

where S = zḡj−sR(wl̄j) and∑n

j=1 S equals zero, since allpositions are referred to their centroids. Thus we are leftwith the first and last term of this expression. The first isindependent from t̄0 while the last cannot be negative.The sum will be evidently minimized when t̄0 = 0, or

t0 = z̄g − sR(w̄l). (7)

That is, the optimal translation is just the difference ofthe global centroid and the scaled and rotated local one.As both centroids are known if given the two sets ofpositions, the optimal translational offset, i.e., t0, can bederived once the scale and rotation factors are found.

4.3 ScalingAt this point, assuming that the optimal translation isgiven as t0 = z̄g−sR(w̄l), we have t̄0 = 0 and hence thesum of squares of the residual errors can be written as

n∑j=1

‖ zḡj − sR(wl̄j) ‖2 . (8)

Expanding the above term to complete the squareform in s and noting that ‖ R(wl̄j) ‖2=‖ wl̄j ‖2, we have

n∑j=1

‖ wl̄j ‖2 [s− F ]2

+

n∑j=1

‖ zḡj ‖2 −

n∑j=1

‖ wl̄j ‖2 F 2, (9)


Trace ITrace ZTrace L

(a) Urban areas (180m×190m)

Trace JTrace STrace W

(b) Suburban areas (245×190m)

Fig. 7. Experiment areas in the New Technology Districtof Wuxi City

where F =∑n

j=1 zḡj ·R(w

l̄j)∑n

j=1‖wl̄j‖2.To minimize the above expres-

sion with respect to scale s, the first square term shouldbe set zero, that is, s = F .

4.4 RotationAt present, the only remaining task is to find the rotationin the plane of global coordinate system. By doing this,final complete solution of the position transformationproblem will be achieved.

The optimal rotation should minimize the sum ofsquares of distances between corresponding points oflocal and global coordinates [14], say, minimize

n∑j=1

‖ zḡj −R(wl̄j) ‖2 . (10)

As the local and global coordinate systems are coplanar,there is an angle between corresponding positions zḡjand wl̄j , denoted as αj . In other words, z

ḡj ·wl̄j =‖ z

ḡj ‖‖

wl̄j ‖ cosαj . Let θ denote the angle the global coordinateshave rotated. The above term can be expanded as followssince the angle αj is reduced by θ.n∑

j=1

‖ zḡj ‖2 +

n∑j=1

‖ wl̄j ‖2 −2n∑

j=1

‖ zḡj ‖‖ wl̄j ‖ cos(αj − θ).

(11)To minimize Eqn. 11, we need to maximize the last term,or A cos θ + B sin θ,where A =

∑nj=1 ‖ z

ḡj ‖‖ wl̄j ‖

cosαj , B =∑n

j=1 ‖ zḡj ‖‖ wl̄j ‖ sinαj .This term achieves

extremum when A sin θ = B cos θ, that is,

θ = arcsin±√

B2

A2 +B2, (12)

one maximizing, and one minimizing Eqn. 10.Accomplishing the coordinate transformation, global

positions are aligned to their corresponding transformedlocal ones, which delineate the structure of true trajectorybetter. In other words, a global position z̄gj is replacedwith w̄gj .

5 EXPERIMENTSTo evaluate the proposed approach, we implement aprototype system of GloCal on the increasingly popular

Android platform and collected data in both urbanand suburban areas. In this section, we first detail theexperiment environments and methodology, followed bythe evaluation of each components. Finally, the overallperformance of GloCal is presented.

5.1 Experiment MethodologyWe implemented GloCal on Android OS using GoogleNexus S phones, which are equipped with accelerom-eters, gyroscopes, and compasses, and as well supportGPS functions. The accelerometers and gyroscopes areamenable of a respective frequency of around 50Hz and800Hz, while the GPS unit can report new data oncelocations change (around a period of 1 second). In theprototypal GloCal, we set the sampling rate of all sensorsand GPS to be as high as possible to record redundant in-formation for the purpose of comprehensive evaluationand analysis.

Our experimental environments are twofold: a built-up urban region around an academic building (Fig. 7a)and a spacious suburban area (Fig. 7b). Generally, rawGPS exhibits different accuracy in these two areas due todistinct natural environments. Trajectories are collectedfrom users automatically when they are walking nat-urally and using their mobile phones for navigation.Each trajectory contains a sequence of global positioningreports and a series of sensor records. All raw sensordata are first sanitized with a lowpass filter for furtheruses, while accelerometer readings are additionally com-pensated for gravity.

Note that no extra behavior constraints are exertedto users for data collection. To obtain the ground truthgeographical positions of the paths user traveled forevaluation, however, users have to walk along our pre-defined paths depicted on a map. The real positioninformation can then be acquired by carefully puttingthe routes into handy digital map services, for instance,Google Maps, as shown in Fig. 7a and Fig. 7b.

5.2 Performance Evaluation5.2.1 Local Positioning PerformanceWe first evaluate the performance of local positioning,i.e., user trajectory delineation based on dead reckoning,and validate the underpinning that the local positioningpreserves the structure of the ground truth picturesquely.

Step Counting Accuracy. We test the FSM based stepcounting algorithm on 3 users by collecting 8 tracesfrom their natural walking with various lengths rangingfrom 10 steps to 300 steps, which are counted by theusers themselves and used as ground truth. Integratingthe results from 24 traces, we inspect the impact of thethreshold (negPeak and posPeak in Fig. 3) in Fig. 8a.Obviously, more than 95% traces are counted precisely ofless than 5 step error when using peak values within anappropriate range, from 1.0 to 1.4 in our experiments.Furthermore, we compare the performance of the pro-posed algorithm with previous methods, including the


0 5 10 15 20 25 30 350

0.2

0.4

0.6

0.8

1

Step Couting Error

CD

F

peak=1.6peak=1.4peak=1.2peak=1.0

(a) Step count accuracy with various thresholds

0 2 4 6 8 10 12 140

0.2

0.4

0.6

0.8

1

Step count error

CD

F

Threshold−basedPeak−basedFSM−based

(b) Comparison with traditional methods

Fig. 8. Step counting accuracy

0 10 20 30 400

0.2

0.4

0.6

0.8

1

Stride Estimation Error (cm)

CD

F

70 samples of 20 users40 samples of 20 users40 samples of 10 users

Fig. 9. Stride estimation accuracywith different size of training data

threshold-based [8] and peak-based [10], on 90 traces ofdifferent lengths from 9 users. As shown in Fig. 8b, theproposed FSM-based approach consistently outperformstraditional methods. The improvements are achieved byelaborately describing a series of state transitions duringhuman footsteps, while previous methods can be easilyaffected by single abnormal sensor measurement. Addi-tion to the robustness to noisy sensor and arbitrary users,another excellent property of the proposed algorithmlays on its irrelevance to the number of steps, thusmingling no accumulative error concerns.

Stride Length Estimation Accuracy. We collect train-ing data from totally 20 training users with variousheights and weights. All users are asked to walk alongtwo pre-defined paths, one with length of 20m and theother of 30m, both in a normal manner. One user’s actualstride lengths (in different traces) are measured as thequotients of the path length to the manually countedsteps he took. We train the model using three differentsets of training data and evaluate the estimation accuracyon 15 testing users in each case. As shown in Fig. 9, theestimation accuracy is reasonably high, yet not perfect,and increases with the size of sample data as well asthe number of training users. Nevertheless, the proposedmethod does not rely on large amount of training datasince a small number of training samples can alreadyprovide satisfactory accuracy. For instance, the strideestimation error is only about 8cm using 40 samples from20 users though it can further decreases to about 6cmwith more samples. In addition, it should be pointedout that slight errors in stride length estimation, albeit doexist, can be gracefully tolerated since a scaling factor hasbeen taken into account in the coordinate transformation.

Dead Reckoning Performance. Concerning thatwhether local positioning could produce precise depic-tion of users’ real trajectories, we now fuse the outcomeof step counting, stride estimation, and direction reckon-ing to illustrate an intuitively qualitative picture of thelocal coordinates of 6 user traces with different shapes.As shown in Fig. 10, it can be perceptively seen that thestructure of most user traces can be precisely resuscitatedby local measurements. Although some trajectories doesnot necessarily match the ground truths precisely, e.g.,

case as shown in Fig. 10b, they still perform muchbetter than the global traces. Such results confirm ourbasic postulation that locally dead-reckoned trajectoriespreserve the truthful structure better than the global GPSmeasurements and thus guarantee the correctness andeffectiveness of the proposed approach.

5.2.2 Positioning Accuracy

Now we turn to the accuracy improvement ofGloCal over GPS. We first evaluate the accuracy ofraw GPS with commodity mobile phones. With ourexperimental phones, the average location error from 14measurements over one week in urban areas is 5m∼8m.In the following, we inspect various factors that mightinfluence the performance of GloCal. Briefly, three pa-rameters are given our attention: the point number ninvolved in the coordinate system, the unit distance dbetween adjacent sample points on a trajectory, and theshapes of trajectories.

Impact of point number. To see impact of n, weset n range from 5 to 100 for a long trajectory andintegrate the results on all scenarios. Fig. 11 illustratethe effect of n with d=1.5m, 2.5m, and 3.5m, respectively.Obviously, on all unit distances, the positioning errorsare significantly reduced with the increase of the numberof points used. Such impressive results are natural sincethat more points forms longer trajectory and thus theinfluence of those strong misaligned GPS measurementsis avoided to a certain extent.

Impact of unit distance. From Fig. 11, one can moreor less see that the positioning accuracy also increaseswhen using longer unit distance d. To further validatethis point, we examine the accuracy improvements onvarious d with fixed n. As depicted in Fig. 12, the resultspan out as we expected that positioning error doesdecrease when d lengthens. On one hand, we suspectthat such results benefit from the better transformationresidual errors under larger unit distances, which re-sulting sparser sample points and thus relaxed structureconstraints. On the other hand, with an identical pointnumber, larger unit distances means longer (but notexcessively long) trajectories, which produces superiorperformance. Albeit counter-intuitive, this crucial prop-


−20 −10 0 10 20 30 40 500

50

100

150

200

X (m)

Y (

m)

True PositionsRaw Global PositionsTransformed Local Positions

(a) Trace I in urban areas

0 20 40 60 80 1000

10

20

30

40

50

60

70

80

X (m)

Y (

m)


(b) Trace Z in urban areas

10 20 30 40 500

20

40

60

80

100

120

140

X (m)

Y (

m)


(c) Trace S in suburban areas

Fig. 10. Local-generated trajectories preserve the structures of ground truth paths precisely.

erty strengthens GloCal’s feasibility and practicabilityas the performance can be consistently guaranteed evenwith low GPS sample frequency (which correspondinglymeans large unit distances).

Impact of trajectory shape and raw GPS distribu-tion. Apart from the two key parameters, i.e., pointnumber n and unit distance d, we also observe thataccuracy of GloCal has no noticeable relevance to thestructure of trajectories, yet is directly influenced by thedistribution of raw global measurements. As portrayedin Fig. 13, GloCal achieves significant improvement onvarious trajectories. However, we observe that when rawGPS measurements deviate the true positions heavily,the accuracy GloCal could achieve will also be limited,Nevertheless, one can always see that GloCal improvesthe global accuracy under all experimental scenarios,which demonstrates its effectiveness in practical usage.

Fusing all results together, we plot the respectiveaccuracy of GloCal in urban and suburban areas inFig. 14a and Fig. 14b, and incorporate them to derivethe overall accuracy in Fig. 14c. All results show thatan impressive improvement of 20%∼30% over raw GPSis achieved while the average error is limited under4m. This accuracy also outperforms the dead-reckoning-only method [8], which provides an average accuracyof 11m in urban regions. We believe GloCal sets up anunconventional perspective and provides a practical wayto improve GPS accuracy using mobile phone only, withnegligible extra energy consumption compared to theGPS only mode.

6 RELATED WORK6.1 Global Positioning TechnologyGlobal positioning technologies, like GPS, GLONASS,and Galileo, have revolutionized a range of location-awareness services with their global coverage and out-standing performance [16]. However, many applicationsstill suffer from global positioning errors due to variousfactors [1]. For the dominant GPS, several augmentationsystems are developed to provide accuracy, availability,or any other improvement. AGPS [3] assists GPS bygaining information via a wireless network, such as theGPS receivers on cell towers which have been accurately

located, to relay the satellite information to the receiver.DGPS [4] looks for differences between the satellite-located positions and the known fixed positions, andbroadcasts such differences to the receivers to providebetter accuracy. A most recent GPS augmented systemis the Wide Area Augmentation System (WAAS) [5],a satellite-based augmentation system operated by theFederal Aviation Administration (FAA), which supportsaircraft navigation across North America. Other aug-mentation systems include IGS, CORS, LAAS, etc [2].Either relying on fixed reference stations with exactlyknown locations, or requiring constant network connec-tions, all these augmentation systems need to be run byspecial operators and are available only in limited areas.

On the other hand, considering problems with GPSbeyond accuracy, including poor indoor supports, largebattery consumption, and long acquisition time, inno-vated algorithms and alternative solutions to globalpositioning are proposed. QuickSync [17] presents a fastGPS synchronization algorithm taking advances in iFFT.Leveraging publicly available information such as GNSSsatellite ephemeris and an Earth elevation database, CO-GPS [18] allows a mobile devices to obtain good qualityGPS locations from a few milliseconds of raw GPSsignals by postprocessing in cloud. Place Lab [19] usesGSM and WiFi signals as fingerprints for localization.Active Campus [20] adopts an idea similar to Place Lab,but assumes that locations of WiFi access points areavailable a prior. Taking advantage of the millions of WiFiaccess points throughout populated areas, Skyhook [21]developed a location system for localization indoors andin urban areas, as a supplementary to GPS. While onlyproviding coarse-grained location information, from tensto hundreds of meters, GSM/WiFi based positioningtechnologies also require war-driving in the target areasto acquire GPS coordinates corresponding to GSM/WiFifingerprints. In addition, these works all aim at supple-menting the coverage of GPS, instead of improving theoriginal GPS accuracy.

6.2 Mobile Phone Localization

The mobile phone localization literature is indeed vast.In the space of interests, we only review the most


0 20 40 60 801

2

3

4

5

6

7

Number of Points

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(a) d = 1.5m

0 10 20 30 40 50 60 70

2

3

4

5

6

Number of Points

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(b) d = 2.5m

0 10 20 30 40 502

2.5

3

3.5

4

4.5

5

5.5

6

Number of Points

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(c) d = 3.5m

Fig. 11. Positioning error decreases with larger numbers of points used.

1 1.5 2 2.5 3 3.5 42

2.5

3

3.5

4

4.5

5

5.5

Unit distance (m)

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(a) n = 40

1 1.5 2 2.5 3 3.52

2.5

3

3.5

4

4.5

Unit distance (m)

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(b) n = 50

1 1.5 2 2.5 31.5

2

2.5

3

3.5

4

4.5

5

Unit distance (m)

Pos

ition

ing

Err

or (

m)

GlocalRaw GPS

(c) n = 60

Fig. 12. Positioning accuracy increases with larger unit distances.

0 5 10 15 20 250

0.2

0.4

0.6

0.8

1

Positioning Error (m)

CD

F

GlocalRaw GPS

(a) Trace I in urban areas

0 5 10 150

0.2

0.4

0.6

0.8

1


CD

F

GlocalRaw GPS

(b) Trace Z in urban areas

0 2 4 6 8 10 12 140

0.2

0.4

0.6

0.8

1


CD

F

GlocalRaw GPS

(c) Trace S in suburban areas

Fig. 13. Positioning accuracy in crowded urban areas and spacious suburban areas.

relevant and representative prior work. In particular, wesurvey the works attempting to leverage inertial sensingto characterize human mobility for localization.

Adhere to the thinking of marine or air navigation,known for centuries, smartphone-enabled dead reckon-ing is well-studied for both indoor and outdoor localiza-tion. [9] combines a foot-mounted inertial unit, a detailedbuilding model, and a particle filter to provide absolutepositioning. Considering the energy issues, GAC [22],CompAcc [8], and WheelLoc [23] all provide localiza-tion in outdoor environments, depending mainly on theaccelerometer and compass sensors and using the GPSinfrequently for initialization and recalibration. Concern-ing GPS is unavailable indoors, recent work Unloc [10]identifies indoor landmarks with unique WiFi (and mag-netic or accelerometer) signatures in an unsupervised

way for zero-calibration localization, while Zee [7] tooenables zero-effort indoor localization by placing dead-reckoned user paths into an indoor map, according tothe constraints imposed by the map. Considering humanmobility together with radio fingerprint space, LiFS [6]successfully releases the site survey process of tradition-al indoor localization by applying human motions toconnect previously independent radio fingerprints andconstruct a fingerprint space.

GloCal differs from previous works towards mobilephone localization in two folds: Firstly, GloCal aims atimproving the GPS accuracy by using inertial sensing asa second, local localization while most previous worksleverage dead-reckoning as an alternative to GPS toreduce the energy consumption and extend the servicecoverage. Secondly, GloCal uses dead-reckoning in a


0 5 10 15 20 250

0.2

0.4

0.6

0.8

1

← mean error =3.9m



CD

F

GlocalRaw GPS

(a) Overall accuracy in urban areas

0 5 10 15 200

0.2

0.4

0.6

0.8

1




CD

F

GlocalRaw GPS

(b) Overall accuracy in suburban areas

0 5 10 15 20 250

0.2

0.4

0.6

0.8

1




CD

F

GlocalRaw GPS

(c) Overall accuracy of GloCal

Fig. 14. Overall positioning accuracy

diametrically different way compared to previous works.Previous works require additional reference information,such as GPS [8], indoor landmarks [10], and digital floorplans [7], to initialize and recalibrate the dead-reckonedpositions. In contrast, desiring solely the relative trajec-tories, GloCal is free of using any reference informationor extra infrastructure and thus is more practical andfeasible in real-world applications.

7 CONCLUSIONIn this paper, we propose an innovative approach to im-prove global positioning accuracy using user trajectoriesmeasured by commodity mobile phones, without anydependence on either fixed infrastructure or additionalreference information. We design a novel smartphone-enabled dead reckoning technique, including step count-ing, stride estimation, and direction reckoning, to ac-curately delineate users’ locomotion. On this basis, theglobal positioning accuracy is refined by fitting the lessaccurate global positions to the structure of the more pre-cise local trajectories. The preliminary experiment resultsin urban and suburban areas suggest that GloCal canachieve 30% improvement on GPS average accuracy,demonstrating its promise in real-world feasibility. Ourongoing work focuses on pursuing GloCal to assist GPSpositioning for vehicles and unmanned aircrafts.

ACKNOWLEDGMENTThis work is supported in part by the NSFC Ma-jor Program under grant No. 61190110, National Ba-sic Research Program of China (973) under grant No.2012CB316200, and NSFC under grant No. 61171067,61133016, 61272429, and 61303211.

REFERENCES[1] E. Kaplan and C. Hegarty, Understanding GPS: principles and

applications. Artech House Publishers, 2006.[2] “Goverment information about the global positioning system

(gps),” http://www.gps.gov/.[3] J. LaMance, J. DeSalas, and J. Jarvinen, “Assisted gps: a low-

infrastructure approach,” GPS World, vol. 13, no. 3, pp. 46–51,2002.

[4] B. Parkinson and P. Enge, “Differential gps,” Global PositioningSystem: Theory and applications., vol. 2, pp. 3–50, 1996.

[5] P. Enge and A. Van Dierendonck, “Wide area augmentationsystem,” Progress in Astronautics and Aeronautics, vol. 164, pp. 117–142, 1996.

[6] Z. Yang, C. Wu, and Y. Liu, “Locating in fingerprint space:Wireless indoor localization with little human intervention,” inProceedings of ACM MobiCom, 2012, pp. 269–280.

[7] A. Rai, R. Sen, K. K. Chintalapudi, and V. Padmanabhan, “Zee:Zero-effort crowdsourcing for indoor localization,” in Proceedingsof ACM MobiCom, 2012, pp. 293–304.

[8] I. Constandache, R. R. Choudhury, and I. Rhee, “Towards mobilephone localization without war-driving,” in Proceedings of the IEEEINFOCOM, 2010, pp. 1–9.

[9] O. Woodman and R. Harle, “Pedestrian localisation for indoorenvironments,” in Proceedings of ACM UbiComp, 2008.

[10] H. Wang, S. Sen, A. Elgohary, M. Farid, M. Youssef, and R. R.Choudhury, “No need to war-drive: unsupervised indoor local-ization,” in Proceedings of ACM MobiSys, 2012, pp. 197–210.

[11] J. Kim, H. Jang, D. Hwang, and C. Park, “A step, stride andheading determination for the pedestrian navigation system,”Journal of Global Positioning Systems, vol. 3, no. 1-2, pp. 273–279,2004.

[12] G. Shen, Z. Chen, P. Zhang, T. Moscibroda, and Y. Zhang, “Walkie-markie: indoor pathway mapping made easy,” in Proceedings ofUSENIX NSDI, 2013, pp. 85–98.

[13] C. Karney, “Transverse mercator with an accuracy of a fewnanometers,” Journal of Geodesy, vol. 85, no. 8, pp. 475–485, 2011.

[14] B. K. P. Horn, “Closed-form solution of absolute orientation usingunit quaternions,” Journal of the Optical Society of America A, vol. 4,no. 4, pp. 629–642, 1987.

[15] J. McGlone, E. Mikhail, J. Bethel, R. Mullen, A. S. for Photogram-metry, and R. Sensing, Manual of photogrammetry. AmericanSociety for Photogrammetry and Remote Sensing, 2004.

[16] B. Hofmann-Wellenhof, H. Lichtenegger, and E. Wasle, GNSS–global navigation satellite systems: GPS, GLONASS, Galileo, and more.Springer Verlag Wien, 2008.

[17] H. Hassanieh, F. Adib, D. Katabi, and P. Indyk, “Faster gps via thesparse fourier transform,” in Proceedings of ACM MobiCom, 2012,pp. 353–364.

[18] J. Liu, B. Priyantha, T. Hart, H. Ramos, A. A. Loureiro, andQ. Wang, “Energy efficient gps sensing with cloud offloading,”in Proceedings of ACM SenSys, 2012.

[19] Y.-C. Cheng, Y. Chawathe, A. LaMarca, and J. Krumm, “Accu-racy characterization for metropolitan-scale wi-fi localization,” inProceedings of ACM MobiSys, 2005, pp. 233–245.

[20] W. G. Griswold, P. Shanahan, S. W. Brown, R. Boyer, M. Ratto,R. B. Shapiro, and T. M. Truong, “ActiveCampus: experiments incommunity-oriented ubiquitous computing,” Computer, vol. 37,no. 10, pp. 73–81, 2004.

[21] “Skyhook: The worldwide leader in location positioning, contextand intelligence.” http://www.skyhookwireless.com/.

[22] M. Youssef, M. Yosef, and M. El-Derini, “Gac: Energy-efficient hy-brid gps-accelerometer-compass gsm localization,” in Proceedingsof the GLOBECOM, 2010, pp. 1–5.

[23] H. Wang, Z. Wang, G. Shen, F. Li, S. Han, and F. Zhao, “Wheelloc:Enabling continuous location service on mobile phone for outdoorscenarios,” 2013.


Chenshu Wu received his B.E. degree in Schoolof Software from Tsinghua University, Beijing,China, in 2010. He is now a Ph.D. student inDepartment of Computer Science and Technol-ogy, Tsinghua University. He is a member ofTsinghua National Lab for Information Scienceand Technology. His research interests includewireless ad-hoc/sensor networks and pervasivecomputing. He is a student member of the IEEEand the ACM.

Zheng Yang received a B.E. degree in computerscience from Tsinghua University in 2006 anda Ph.D. degree in computer science from HongKong University of Science and Technology in2010. He is currently an assistant professor inTsinghua University. His main research interest-s include wireless ad-hoc/sensor networks andmobile computing. He is a member of the IEEEand the ACM.

Yu Xu received his B.E. degree in automationand Ph. D degree in control science and en-gineering from Zhejiang University in 2003 and2008. He is currently a lecturer in Wenzhou Uni-versity. His research interestes include robotics,embedded systems and wireless sensor net-works.

Yiyang Zhao received the B.S. degree fromTsinghua University, the Mphil degree from theInstitute of Electrical Engineering of CAS, andthe PhD degree in computer science from theHong Kong University of Science and Technol-ogy, in 1998, 2001, and 2010, respectively. Hisresearch interests include RFID, pervasive com-puting, distributed systems, embedded systems,and high-speed networking. He is a member ofthe IEEE and IEEE Computer Society.

Yunhao Liu received the BS degree in automa-tion from Tsinghua University, China, in 1995,the MS and PhD degrees in computer scienceand engineering from Michigan State University,in 2003 and 2004, respectively. He is now EMCChair Professor at Tsinghua University, as wellas a faculty member with the Hong Kong Uni-versity of Science and Technology. His researchinterests include wireless sensor network, peer-to-peer computing, and pervasive computing.He is a senior member of the IEEE.

Documents

TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, … · Human Mobility Enhances Global Positioning Accuracy for Mobile Phone Localization Chenshu Wu, Student Member, IEEE, Zheng