ADVANCED PEDOMETER FOR SMARTPHONE-BASED

ACTIVITY TRACKING

Michal Tomlein, Pavol Bielik, Peter Kr´atky,

Stefan Mitr´ık, Michal Barla and M´aria Bielikov´a

Institute of Informatics and Software Engineering, Faculty of Informatics and Information Technologies

Slovak University of Technology in Bratislava, Ilkoviˇcova 3, 842 16 Bratislava, Slovakia

Keywords:

Advanced pedometer, Activity tracking, Cheating detection, Neural networks.

Abstract:

This paper describes the design of an advanced smartphone pedometer developed as part of a larger solution

aimed at encouraging a healthier lifestyle through regular physical activity, called Move2Play. Move2Play

provides several motivational methods to promote exercise, some of which are based on the use of points

derived from tracked activity. While current pedometers are easily tricked into counting steps by shaking the

device, our advanced pedometer uses a neural network to detect and prevent this kind of cheating. In this paper,

we discuss both the method for counting steps and our innovative approach to the recognition of cheating.

1 INTRODUCTION AND

RELATED WORK

Evolution in IT has resulted in signiﬁcant changes in

our everyday life. Information has become ubiqui-

tous, can be accessed almost anywhere and anytime

and be leveraged to deal with our tasks more efﬁ-

ciently. Moreover, we now have access to new kinds

of information previously unavailable. One such kind

of information that can have a major impact is an

overview of our everyday physical activity. Having

this information is a basic prerequisite if we want to

tackle one of the biggest problems of our society – the

lack of physical activity, overweight and obesity.

This information can be captured and interpreted

by an intelligent wearable device. However, the re-

quirement to wear an additional measuring device is a

key barrier to achieving long-term sustainability (Fu-

jiki, 2008). It is therefore preferable to rely on a de-

vice people already carry, i.e. a smartphone.

Currently available solutions track users’ move-

ment “on-demand”, mainly using GPS sensors. How-

ever, they are mainly used by already rather physi-

cally active people, while most of the people who

need to become more physically active are unaware

of such solutions or are not motivated to use them

as they do not exercise regularly. We focus on such

people who require additional motivation in order to

become more physically active.

There is an activity we perform naturally, with-

out paying much attention to it – walking. Walking

reduces rates of chronic disease, ameliorates health

care costs (Lee and Buchner, 2008) and is one of a

few activities that every age group can and should

participate in. However, over the last few decades,

the amount of walking people do every day has been

on the decrease, which has been one of the reasons

of increased overweight and obesity and other health

problems.

To change this, we track and analyse walking

throughout the day, not “on-demand” but continu-

ously, using smartphone sensors. Once captured, we

provide users with a visual overview of their perfor-

mance, which is the ﬁrst step towards improvement.

However, activity tracking and visualisation alone are

insufﬁcient, as the drop out rate is a serious problem

even in professional training programmes and can be

as high as 89 % (Iwane, 2000).

Therefore, we proposed a broader concept of com-

prehensive physical activity management, driven by

various motivational factors supporting its users in

achieving the required amount of daily activity. The

process consists of recommendation of the appropri-

ate amount of activity in the form of a personalised

daily plan and automatic tracking and evaluation of

performed activity. An important part of the process is

the integration of different motivational factors, both

intrinsic and extrinsic, which ensure its long-term sus-

tainability. We implemented this process in a solu-

tion called Move2Play (Bielik, 2012). This paper de-

scribes the design of an important part of our solution

– an advanced pedometer.

Most current smartphones come with an ac-

401

Tomlein M., Bielik P., Krátky P., Mitrík Š., Barla M. and Bieliková M..

ADVANCED PEDOMETER FOR SMARTPHONE-BASED ACTIVITY TRACKING.

DOI: 10.5220/0003786704010404

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 401-404

ISBN: 978-989-8425-88-1

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

celerometer, which is widely used by pedome-

ter applications for step counting. According

to (Marschollek et al., 2008), there exist only a

few well-known freely available accelerometer-based

step detection algorithms. Most of these approaches

rely on threshold-based detection and they generally

achieve sufﬁcient precision.

Motivating people to be more physically active is

a non-trivial task. In our case, we use measured ac-

tivity as an input to motivational mechanisms such as

various kinds of rewards. Especially with children,

there is the risk that they will choose the easier path

to get to the rewards offered, which means cheating,

i.e., imitation of steps by shaking the phone in hand,

becomes a problem. Current pedometers, both dedi-

cated devices and phone applications, do not address

this problem as they rely on motivated users.

2 STEP DETECTION BASED ON

SMARTPHONE SENSORS

The accelerometer in a mobile phone sends events in

regular intervals with the three components of accel-

eration (x, y, z), which range between -1 and 1. Be-

cause the orientation of the phone inside the pocket

is not known in advance and may be different each

time, it is convenient to track the changes in the di-

rection of the vector of acceleration. For this purpose

we use the dot product of two consecutive accelera-

tions. Given normalised vectors, this is the cosine of

the angle between them – let it be d (1). When the

phone stops abruptly and reverses direction, this an-

gle should equal 180

◦

. In practice, this is rarely the

case, since the phone slips inside the pocket.

= (cosα)

i−1

+ y

i−1

+ z

i−1

+ y

+ z

i−1

+ y

i−1

+ z

i−1

(1)

During inactivity and between successive steps, d

equals 1. A negative peak (a decrease below a thresh-

old) is considered a step. This threshold is different

for the left and right leg, depending on which is closer

to the phone. To increase the precision of the mea-

surement and to ﬁlter out noise, we use a weighted

moving average of the last 10 values of d (2).

WMA

(d)

10d

+ 9d

i−1

+ · ··+ d

i−9

(2)

Figure 1 shows 20 walking and running steps. In

case of walking, it is straightforward to recognise

steps taken with the right foot from steps taken with

the left foot. In case of running, the difference is not

sufﬁciently recognisable.

WMA

(cos )

501 451 401 351 301 251 201 151 101 51 1

0.2

0.4

0.6

0.8

1.2

iteration

Figure 1: Walking (dashed line) and running (solid line) –

20 steps (LG with Windows Phone 7).

Figure 2 shows two forms of cheating, slow and

fast shaking of the phone in hand, in contrast to walk-

ing. Similarly to previous ﬁgures, the lines denote

the weighted moving averages of d values as given

by (2). Greater regularity can be seen in both forms

of cheating than in walking or running, as well as

more frequent alternation of minima and maxima.

The weighted moving average often does not reach

the value of 1, as is the case in between walking and

running steps.

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.1

1 51 101 151 201 251 301 351 401 451 501

iteration

Cheating (slow)

Cheating (fast)

Walking

WMA

(cos )

Figure 2: Cheating by slow and fast shaking of the phone in

contrast to walking (iPhone).

Our algorithm utilises threshold-based detection

of steps with different thresholds for left and right

legs. When the weighed moving average as given by

(2) decreases below these thresholds, the algorithm

increments the step count. The thresholds are de-

ﬁned differently for various models of smartphones

and are based on experimental results. The values for

the particular devices are shown in table 1. In the next

stages of development, the Wolf method may be used

to eliminate the hardware-speciﬁc constants.

The algorithm takes several attributes of walking

into account:

• the alternation of the left and right legs – via spe-

ciﬁc thresholds,

• vibrations after the foot strikes the ground – using:

– minimum step duration,

– minimum required deceleration computed for

a given value of d before acceleration can take

place.

HEALTHINF 2012 - International Conference on Health Informatics

402

The core logic of our algorithm without the above

considerations every hundredth of a second executes

the following steps:

1. Calculate the current value of d using (1) and av-

erage using (2).

2. If the average is lower than the threshold for the

right (left) leg: increase the step count.

Table 1: Thresholds for step detection.

iPhone HTC HD7

Samsung Omnia 7

LG Optimus 7

Right leg 0.96 0.97 0.9

Left leg 0.98 0.995 0.985

3 INTELLIGENT DETECTION OF

CHEATING

Figure 2 shows the striking difference between walk-

ing and cheating, which can be described as a differ-

ence of patterns. Searching for patterns in measured

data can be done in several different ways. One solu-

tion is the creation of a decision tree. However, un-

like a decision tree designed to simply count steps,

a tree capable of detecting those patterns of cheating

that make it different from walking or running, would

necessarily be considerably more complicated.

A common solution to problems which involve a

search for patterns in measured data is the training and

application of an artiﬁcial neural network. This is the

approach we explore and test the applicability of.

Given that the solution is to be deployed on a mo-

bile platform, where computational complexity is un-

desirable because of battery life constraints, we chose

a feedforward neural network with one hidden layer.

We opted for 50 inputneurons, which is the number of

weighted averages of d calculated each second. It is

also the approximate number of averages necessary to

distinguish between cheating and valid activity using

the unaided eye, the patterns being clearly visible. We

need only a single output neuron, whose output value

we deﬁned as 0 for valid activity and 1 for cheating.

We prepared a total of 269 training sets of 50 inputs.

Following an analysis of the input sets, we chose the

incremental training method as the most suitable one.

We found that the sigmoid function is the most suit-

able for the hidden layer, whereas the linear symmet-

ric function was chosen for the output layer.

We implemented the neural network in our appli-

cation for the iOS platform, i.e., for the iPhone. We

used the FANN Library

, one of the best-known and

most widely deployed libraries for training and the

application of artiﬁcial neural networks.

In order to evaluate the neural network, we cre-

ated separate sets of measured data for walking, run-

ning and cheating. Figure 3 shows typical outputs

of the neural network for walking. When determin-

ing whether or not the user is currently cheating, a

weighted moving average of the last 4 outputs of the

neural network is used. As can be seen, the average

rarely exceeds the value of 0.5.

-1.5

-1

-0.5

0.5

1.5

1 21 41 61 81 101 121

iteration

Expected

output

Network

output

Moving

average

WMA

(cos )

Figure 3: Outputs of the neural network while walking.

Figure 4 shows typical outputs for running. As

can be seen, the neural network has a higher tendency

to classify running as cheating compared to walking.

In spite of that, the weighted moving average of the

last four outputs rarely exceeds the value of 0.7.

-1.5

-1

-0.5

0.5

1.5

1 21 41 61 81 101 121

iteration

WMA

(cos )

Expected

output

Network

output

Moving

average

Figure 4: Outputs of the neural network while running.

Figure 5 shows typical outputs of the neural net-

work for cheating, which signiﬁcantly differ from the

previous values for walking and running.

-1.5

-1

-0.5

0.5

1.5

1 21 41 61 81 101 121 141 161 181

iteration

WMA

(cos )

Expected

output

Network

output

Moving

average

Figure 5: Outputs of the neural network while cheating.

The most frequent output of the neural network

when cheating is 1. The weighted moving average of

the last four outputs mostly stays above 0.5.

Table 2 shows a comparison of the desired and av-

erage output of the neural network. The results are

The Fast Artiﬁcial Neural Network Library is available

at: http://leenissen.dk/fann/wp/

ADVANCED PEDOMETER FOR SMARTPHONE-BASED ACTIVITY TRACKING

403

sufﬁciently accurate for walking. There is room for

improvement especially in the case of running.

Table 2: A comparison of the desired and average outputs

of the neural network.

Activity

Sample

size

Desired

output

Average

output

Standard

deviation

Walking 693 0 0.0612 0.3175

Running 151 0 0.2116 0.3654

Cheating 251 1 0.818 0.257

Solving the Problem with Punning. At ﬁrst glance

it may seem problematic to distinguish running from

cheating, since the outputs of the neural network fre-

quently overlap. The solution involves the current

minimum of d from (1), which will be lower for run-

ning steps than for walking steps. This is also the case

with fast as opposed to slow shaking of the phone in

hand during cheating.

The problem can be solved using a moving thresh-

old between valid activity and cheating. As an exam-

ple, a threshold of 0.5 can be chosen for values typ-

ically measured when walking or cheating by slower

shaking. As more vigorous activity is detected, the

threshold may be raised to a higher value, such as 0.7.

This is possible as a consequence of the fact

that more intensive shaking sees the neural network

produce outputs of 1 with a higher probability than

slower, less intensive shaking. The moving threshold

can therefore be used without fear of misinterpreta-

tion of cheating for running.

It is important to realise, though, that in the cur-

rent state the neural network fulﬁls the purpose it was

designed for. The purpose of the implementation of

the neural network was not to detect cheating with an

accuracy of a hundred per cent. First and foremost,

the purpose was to discourage cheating by making it

sufﬁciently difﬁcult and tedious, so that it is more re-

warding to take a walk and receive points for real ac-

tivity. The current implementation of our neural net-

work meets this goal.

4 CONCLUSIONS

In this paper, we describe the design and implementa-

tion of an advanced smartphone accelerometer-based

pedometer, which uses an innovative approach to the

detection of cheating. This is important for applica-

tions where activity is a base for rewarding and there

is no natural way of ensuring motivation for being ac-

tive (typically for children).

We have based our solution on the recognition of

patterns characteristic to cheating. To that end, we

have successfully trained and used a feedforward ar-

tiﬁcial neural network. The results we have obtained

in evaluation conﬁrm the applicability of an artiﬁcial

neural network for the detection of cheating. We have

been able to detect cheating with a sufﬁcient precision

and have thus met our goal of discouraging potential

users from such an activity.

We have developed our advanced pedometer as

part of a larger solution. Speciﬁcally, the pedometer is

part of the activity tracking module of our Move2Play

system. The activity tracking module uses GSM sig-

nal strength ﬂuctuation analysis to determine whether

an activity is taking place or not and thus to save bat-

tery life (Bielik, 2011). Activity tracking together

with personalised activity recommendation and eval-

uation are the essential parts of our solution. In order

to evaluate our solution, we realised its specialisation

called Move2PlayKids, which depends on the intelli-

gent pedometer module as children who need to exer-

cise more tend to cheat.

ACKNOWLEDGEMENTS

This work was partially supported by the grants

VG1/0675/11, KEGA 028-025STU-4/2010 and it is

the partial result of the OP R&D for the project ITMS

26240220039, co-funded by the ERDF.

REFERENCES

Bielik, P., e. a. (2012). Move2play: An innovative approach

to encouraging people to be more physically active.

ACM SIGHIT Int. Health Informatics Symposium.

Accepted.

Bielik, P. (2011). Personalized training plan recommenda-

tion and activity tracking for a healthier lifestyle. Bul-

letin of the ACM Slovakia, 3(4).

Fujiki, Y. e. a. (2008). Neat-o-games: blending physical ac-

tivity and fun in the daily routine. Comput. Entertain.,

6:21:1–21:22.

Iwane, M., e. a. (2000). Walking 10,000 steps/day or more

reduces blood pressure and sympathetic nerve activ-

ity in mild essential hypertension. Hypertens Res,

23(6):573–580.

Lee, I. and Buchner, D. M. (2008). The importance of walk-

ing to public health. Medicine & Science in Sports &

Exercise, 40(7).

Marschollek, M. et al. (2008). A performance comparison

of accelerometry-based step detection algorithms on a

large, non-laboratory sample of healthy and mobility-

impaired persons. In Engineering in Medicine and Bi-

ology Society, EMBS 2008, pages 1319 –1322. IEEE.

HEALTHINF 2012 - International Conference on Health Informatics

404