ArticlePDF Available

The Federal Communications Commission and the Broadcasting Industry

Authors:

Abstract and Figures

The proliferation of sensor networks employing wireless data transmission technologies has paved the way for the collection of large amounts of measurement data. Several research teams have used this opportunity to develop algorithms aimed at gaining information from sensor data. Motion detection is one of the most actively researched areas. In this article, we present a system for examining motion detection in a general environment. In other words, motion forms are not identified with various wearable sensors; instead, we use the data collected by the sensors of mobile phones kept with almost all members of society now.
Content may be subject to copyright.
ActaUniversitatisSapientiae
Electrical and Mechanical Engineering, 8 (2016) 29-41
29
Development of Motion Detection Algorithms Based on
Simultaneous Execution Using Mobile Phone Sensors
Zsófia SÁNDOR1, Gergely KIS2
1 Data scientist, research manager, Budapest, e-mail: sandorzsofi@yahoo.com
2 Department of Infocommunication, Corvinus University of Budapest, Budapest,
e-mail: gergely.kis@uni-corvinus.hu
Manuscript received February 15, 2016; revised December 24, 2016
Abstract: The proliferation of sensor networks employing wireless data
transmission technologies has paved the way for the collection of large amounts of
measurement data. Several research teams have used this opportunity to develop
algorithms aimed at gaining information from sensor data. Motion detection is one of
the most actively researched areas. In this article, we present a system for examining
motion detection in a general environment. In other words, motion forms are not
identified with various wearable sensors; instead, we use the data collected by the
sensors of mobile phones kept with almost all members of society now.
Keywords: Motion detection, sensor data.
1. Introduction
A boom of wireless communication technologies and specifically mobile
phones has brought about institutional, social and cultural changes, just as
predicted by the American writer George Gilder [1]. Technological innovation
offers solutions to an increasing number of issues that used to seem insoluble
due to technological constraints and the lack of measured data.
Within technological innovation in general, progress in terms of decreasing
calculation costs has been especially fast. Consequently, tasks can now be
tackled that require the processing of large quantities of information. A current
buzzword is “Big Data”, meaning the handling of data quantities that cannot be
managed with conservative data processing methods and tools. Various sensory
information is a typical source of such high-volume data. The second relevant
direction of technological development is related to sensor development:
sensors are decreasing in size, becoming more and more accurate, and consume
less energy. The spread of wireless sensor networks (WSN) allows for the real-
DOI: 10.1515/auseme-2017-0003
Unauthenticated
Download Date | 12/19/17 10:39 AM
30 Zs. Sándor, G. Kis
time transfer and processing of the data collected by a sensor, and also for
connecting several sensors.
Studying human behavior is a key social research area. Gaining an insight
into human behavior could yield widely applicable knowledge, for example in
the fields of medicine, psychology, economic marketing, and health care. One
of the projections of human behavior is physical movement; observing
movement may lead to conclusions about the individual’s behavior and habits.
Thus, scientists are facing a specific range of issues, but also have the toolset
that could provide answers to those issues. Hence today’s wide-ranging research
is made about the usability of sensor-collected data for motion detection. Most
of that research is conducted in laboratories, with sensors attached to test
subjects’ bodies.
This article presents a system based on the processing of data generated in
everyday life, outside of the laboratory environment. The definition of the
problem at hand is followed by the description of the data collection and
transfer system, by the presentation of processing algorithms built on each
other, and finally by a summary and conclusions.
2. Problem definition
A. Historical research
Sensors built into smartphones and other wearable devices can be used in a
variety of ways. With the right method, they are suitable for detecting
practically all forms of motion, from basic step recognition to recognizing
complex actions and providing motivation for a healthy lifestyle. This variety
(both in terms of methods and areas of utilization) is reflected by international
technical literature.
Some research is aimed at specific areas, e.g. at finding the most accurate
step counting method [2]. Using accelerometers, the margin of error is just a
few steps. In an advanced version of this solution, speed can also be taken into
consideration as a factor [3]. In this research, steps were counted during slow
and fast walking, downhill and uphill walking, and while climbing stairs. The
researchers used gyroscopes because they had found that an accelerometer does
not yield accurate step numbers if the test subject walks slowly. The results are
encouraging: the accuracy of step counting in slow walk was above 96% on a
flat surface, more than 95% on inclining and declining routes, and higher than
90% when climbing stairs.
Basic actions such as walking, running, climbing stairs, sitting, standing,
using an elevator, and jumping have been examined by several research teams.
Some of these achieved 99.97% accuracy in action recognition using the
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 31
Random Forest method [4], while the Nearest Neighbor method yielded 93%
accuracy in [5]. The limitations were only one phone, held in the test subject’s
trouser pocket in [4]; in another case [5], a single phone model was used, but at
two locations.
Resting positions (standing, sitting, lying) can be detected highly accurately
[6]. Besides these, walking and climbing stairs were examined in this research,
which resulted in 86% accuracy using a decision tree. Others went farther than
that, detecting car driving in addition to walking, running, cycling, sitting and
standing in [7]. In the research, the effectiveness of the QDA (Quadratic
Discriminant Analysis) and the k-Nearest Neighbor algorithms was tested in
online and offline mode. In online mode, the accuracy of QDA was 95.8%, and
that of k-NN was 93.9%; offline, QDA yielded 95.4%, and k-NN 94.5%. The
phone position was fixed in these cases, too, and in research no. [7], its
direction was also fixed.
In [8], the goal was to identify actions regardless of the position of the
phones; but in this case, too, only the basic actions were examined.
In certain publications [9, 10, 11], the objective was to present a connection
between a healthy lifestyle and movement detection. In one case [9], the
researchers developed a step counter using accelerometer data and a neural
network, aiming to detect false steps (a common error in step counters). The
ultimate goal was to increase the reliability of the health preservation system. In
another research, algorithms were developed for the recognition of uphill and
downhill walking, walking on a flat track, climbing and descending stairs, and
running in [10]. Again, the goal was health preservation. The accuracy of
recognizing these actions was 93.2%, 97.4%, 97.6%, 98.8%, 92.2%, and 90.8%
respectively. Not all analyses are aimed at the processing of health data: based
on a market approach, a convenience function has been developed, focused on
automatically changing the phone’s settings at the start of certain activities such
as running [11].
Besides the basic actions, the recognition of complex activities (such as hand
washing, house cleaning, cooking etc.) has also been researched in [12]. A wide
range of methods was employed. These (apart from the Naïve Bayes method)
were suitable to identify basic actions with over 90% accuracy. However, the
highest accuracy in the case of complex actions was merely 50%.
In most cases, wearable sensors provide good estimations for the user’s
activities; because of the fixed location [13, 14, 15].
As shown above, research is progressing towards several dimensions of
complexity. Our goal is to separate as many medium-complexity actions as
possible. (These include, for example, riding a bus, but not dish washing.) For
wide-ranging usability, that should be achieved with mobile phones regardless
of model and location.
Unauthenticated
Download Date | 12/19/17 10:39 AM
32 Zs. Sándor, G. Kis
B. The challenge
The objective of the system presented is to accurately identify movement in
its natural environment, without additional devices. A solution may not become
wide-spread if the users find the required device inconvenient or uncomfortable.
One example is the need to use an ankle strap and a wristband at the same time
may force users to change their daily routines. This is why mobile phone
sensors were selected in our solution. People keep their phones with them
during the active part of the day, so the phones usually move together with the
users.
However, mobile phones present challenges other than algorithm
development.
The first challenge is related to eliminating the above-mentioned
inconvenience. Mobile phone usage is convenient if the system meets certain
basic requirements such as low energy usage and data transfers, and optimized
resource (e.g. memory) usage. These requirements must definitely be met for a
solution to be easy to use.
The second group of challenges arises from variances across both users and
devices. Movement detection is difficult because of differences between
people’s physiques, habits and movements; the same actions vary from person
to person if we rely on time-series data of sensors. In addition, differences
between mobile phone makes and models should also be addressed; the quality
and accuracy of built-in sensors vary, and even the measurement units may
differ. These challenges need to be resolved for the wide-spread proliferation of
a system.
The third challenge involves ensuring the right response time with sufficient
accuracy. The system will not be used in a lab environment, so, according to
international literature, we should not expect the models to be as accurate as in a
lab, especially because some of the activities to be differentiated are similar to
each other.
C. Technical solution concept
Before outlining the solution concept, it is important to note that two options
are currently available to researchers concerning the technicalities of data
transmission and processing. Each of these options has its advantages and
disadvantages, as explained below.
The problem is that sensor data are available in the mobile phone or other
sensor-equipped device. They either need to be transferred to a processing
server, or the processing must be performed by the device itself. In the former
case, the issue of high data transfers and thus higher energy consumption
must be resolved as this may be inconvenient to users if the data are transferred
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 33
via a mobile Internet connection. Also, the increased data intensity means that
the server’s capacity must be sufficient. On the other hand, the advantage of this
option is that there is no data loss.
The main argument for in-device processing is that not all the data are
required; algorithms running on the device can be used to obtain material
information, and only that is transferred to a server. This is a major advantage
because it saves both device and server resources. On the other hand, running
the processing algorithms uses more memory, and the running jobs make the
energy consumption higher, which can also be inconvenient to users. The
advantages and disadvantages of the two methods are summarized in Table 1
below.
Table 1: Assessment of data processing methods
Data processing on
server
Data processing on
device
Amount of data to be
transmitted
high
low
Local memory usage
low
high
Server storage space
requirement
high
low
Information loss
no
yes
Energy cost
higher due to the data
transfer
higher due to the
running algorithms
We have opted for data processing by a server, primarily because we strove
to keep all information obtained during the research.
In order to resolve the above issues and tackle the challenges, the system
shown in Fig.1 has been created. As we collect data from mobile phone sensors,
a mobile phone application is one of the central elements of the system. This
application collects data based on the configuration received from the
messaging server, writes those data into a file, compresses the file, and transfers
it to the processing servers in the required format. The role of the data server is
to pre-process the high quantity of received data, and to send them to the
algorithm servers. The processing servers run the algorithms, and then estimate
which activities the user is probably performing. The final decision is made
based on several partial decisions.
Unauthenticated
Download Date | 12/19/17 10:39 AM
34 Zs. Sándor, G. Kis
Figure 1: Data processing
3. Measurement and data collection
Defining the right data collection methodology is essential both for
modelling and maintaining the operability of the system. As customizability as
well as easy and convenient usage are paramount to users, we aimed not only
for collecting and transferring data in the highest quality possible, but also for
optimal resource utilization.
The mobile phone application we have developed has a built-in algorithm to
sense low-activity periods when data are not collected. Naturally, this so-called
sleeping mode is customizable via the configuration of the messaging server,
i.e. the running and discontinuation of measurement can be controlled. The
sleep mode optimizes both data traffic and battery usage and also the energy
consumption.
Sensor data are obtained through the sensor’s API (Application
Programming Interface) and are stored in a manner that minimizes data size as
much as possible. This size reduction consists of two parts: transforming and
encoding the file in an optimized size; and compression. The resulting file
(ready for transfer) is about 12% of the original data size. In addition to this size
reduction, sleep mode may significantly lower the data transfer volume,
depending on the activity ratio. The range of sensors to be used for
measurement can also be configured; the algorithms have been developed with
the accelerometer, GPS and gyroscope in focus.
Transferring the collected data to the server in the manner described above
causes minimal inconvenience to the user. This means that the energy
consumption is reduced, and also the amount of processing on the device itself
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 35
keeps the user experience high. All further work phases of the system are
independent of the user, i.e. they do not influence the user at all.
The data are received by the data server’s pre-processing module which
prepares the data for processing, i.e. decompresses and decodes them, and runs
processes required for subsequent algorithms. These include ensuring the
coherence of the time series, decreasing their dimensions, and interpolation in
order to eliminate inaccuracies (minor differences in sampling at a given
frequency) in the sensor API. One of the gravest problems of data collection
with mobile phones is the variance in the type and quality of sensors built into
the devices of various manufacturers; the measurement scales may differ, and
the accuracy of the sensors almost certainly varies. Figure 2 shows deviations in
accuracy. The y axis is the acceleration in m/s2, and the x axis is the millisecond
of the measurement.
Figure 2: Comparison of the operation of mobile phone sensors
Fig. 2shows time series of the accelerometer sensors of the mobile phones
tested. All phones were stationary, but the sensor readouts indicate different
mean values and standard deviations. These differences must be managed with
algorithms. The figure above shows data from devices running the Android
operating system only; iOS measurements are very different due to the data
being recorded in different measurement units. Android phones measure
acceleration in m/s2, while iOS in g. The gravity in common is measured in
units of acceleration, g means the measurement unit of gravity, thus 1 g equals
to approximately 9.81 m/s2 in our case. So unifying the measurement scales is
an important task within pre-processing phase.
4. Algorithm development
A. Activities
The first step in algorithm development involves defining the activities to be
identified. Several scenarios are examined in the related technical literature
Unauthenticated
Download Date | 12/19/17 10:39 AM
36 Zs. Sándor, G. Kis
using the methods described above, from simple step count to attempting to
identify the placement of the device. In our own research, we attempted to
identify actions regardless of the placement of the user’s phone, i.e. we seek
patterns that are common to all phone placements. As an example, let us take a
phone kept in the user’s trouser pocket or purse. These cases differ because a
phone kept in a trouser pocket moves together with the user’s body, while a
telephone kept in a purse moves partly independently of the body. But the basic
movement pattern must be the same in both cases; it is this pattern that we aim
to find with algorithms, ignoring the noise caused by the placement of the
device.
During the research, we collected as detailed reference data as possible, but
some of the actions were aggregated because they were similar. Reference
measurements for the following actions were made:
Human movement: immobility (phone placed on a desk or kept with the
user); walk; run; ascending and descending stairs; phone usage (calls and
screen usage);
Vehicular movement: bus; metro; tram; trolley; train; suburban train;
Other: lift and escalator (ascending or descending).
B. Concurrent algorithms
After pre-processing of accelerometer data, the system processes branch off
so that several algorithms can run concurrently for the best possible end-result.
As each algorithm has strengths and weaknesses, they would not be sufficient
individually. But they support and improve each other for a robust overall
system.
For example, one algorithm, which is based on GPS usage, can significantly
improve the results of other algorithms. But for that, the user must turn on GPS-
based location identification in the phone’s menu; and even if that is enabled,
there is no GPS connection indoors. So, that algorithm is used as an accessory
function only; our algorithms are mostly based on an accelerometer.
Pre-processing makes the data easily usable for the concurrent algorithms.
This means, that all the data packages that the algorithm servers receive do not
need any additional calculations.
There are two types of algorithms. The first type is when we identify
episodic types of events while the second is when we identify the activities as
macro-processes.
All of the activities have typical episodes. For example, when a vehicle
accelerates; when we take a step; when we sit down and so on. These episodes
last for very short periods but they are pretty easy to identify. The first
algorithm catches these episodes and is based on clustering. The variables of the
clustering are the characteristics of the time series. We used the k-means
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 37
clustering algorithm for generating the static clusters for the episodes, and the
distances from the cluster centers helps us to assign the new samples to the
clusters and recognize the activities this way.
The second group is the macro-process type algorithms. The first amongst
these algorithms is the GPS algorithm. The GPS algorithm is mostly based on
the calculation of the speed based on the time and the distance. GPS coordinates
are pretty accurate, so we can use the haversine formula to calculate the distance
between two coordinate pairs (1).
= 2221
2+12221
2
Where r is the radius of the Earth that is approximately 6371km, λ2- λ1 is
the difference of the two longitudes and ϕ2-ϕ1 is the difference of the two
latitudes. The computed and corrected speed values help us to identify the
activities, as the activities can be well separated using these values.
The second algorithm of this group is based on the similarity of the 3-
dimension accelerometer time series. Time series data points can be paired to
each other and as acceleration is measured in 3-dimensions Euclidean
distance can be calculated in 3 dimensions (2). Using the distances, activities
can be categorized by comparing the differences to the reference data.
,= ()2
=1
=(11)2+ (22)2+ (33)2 (2)
The third algorithm of this group is based on the amplitude of the time
series. This algorithm is the best in recognizing the real macro activities. This is
able to recognize the stops of a public transport vehicle and this way it can
identify the travelling on public transportation as a process. The highest and
lowest amplitude sections’ patterns are unique by the activities.
The last algorithm in this group is based on a matrix model. In this model,
we create unique “masks” of each reference activities. Activities differ in the
sense how data points follow each other so if we use a matrix to identify the
sequentiality of the data points, we can create matrices that can be used to
identify the unknown activities. These “masks” create a similarity measure that
varies between -1 and 1, and where 1 means the identity.
C. Real life examples
Table 2 and 3 each show the similarities/differences between the activities
for one algorithm. The figures indicate the similarity of the selected basic
actions.
Unauthenticated
Download Date | 12/19/17 10:39 AM
38 Zs. Sándor, G. Kis
Table 2: Similarities of basic actions as indicated by sensor data
Table 2 contains the similarity measures of the matrices that we have created
in the matrix model. The theoretical minimum value is -1, while the theoretical
maximum is 1. The main diagonal indicates the average similarity between
identical actions, while the rest shows the distance between the actions. Note the
strong similarity between walking and stair climbing, which is caused by the
fact that stair climbing consists of steps, too. Interestingly, the differences
between users’ movements are so big that the two actions can hardly be
distinguished. The only difference between them is in their amplitude, which
this algorithm is less suitable to detect. Note the similarity of moving in an
elevator and being immobile; the reason is that a lift usually moves in a straight
line at a steady pace apart from the initial acceleration and the concluding
deceleration, i.e. an accelerometer indicates the same data as in the case of
immobility. Acceleration is zero both in case of immobility and during straight
movement at a steady speed.
Table 3 shows Euclidean distances in the 3-dimension accelerometer time
series measured in m/s2 between reference activities hence the zeros in the
main diagonal. Some activities constitute groups of stronger similarity; and
walking is similar to climbing stairs in this case as well. Also, the data
measured in vehicles are similar. But this algorithm seems to better distinguish
riding an elevator from a state of immobility, so it is important to combine
several algorithms.
bus stairs lift metro immobility walking
phone
usage
tram train
bus 0,91 0,67 0,25 0,70 0,31 0,66 0,71 0,54 0,65
stairs 0,67 0,95 0,36 0,46 0,51 0,95 0,43 0,39 0,46
lift 0,25 0,36 0,89 0,45 0,77 0,40 0,06 0,62 0,43
metro 0,70 0,46 0,45 0,77 0,44 0,47 0,58 0,66 0,70
immobility 0,31 0,51 0,77 0,44 0,82 0,55 0,12 0,56 0,43
walking 0,66 0,95 0,40 0,47 0,55 0,97 0,43 0,41 0,48
phone
usage
0,71 0,43 0,06 0,58 0,12 0,43 0,80 0,38 0,59
tram 0,54 0,39 0,62 0,66 0,56 0,41 0,38 0,70 0,63
train 0,65 0,46 0,43 0,70 0,43 0,48 0,59 0,63 0,82
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 39
Table 3: Distances between reference actions
For robustness, development was supported by large-sample data collection
in order to map differences between users and to avoid “over teaching”, a
frequent error when developing models. In this modeling case, it would be
rather problematic if the algorithm learned too much from a single user’s
movement and used that to draw conclusions about the movement of other
people who walk slower or use a different means of public transport the patterns
of which cannot be observed in the first user’s case, so the algorithm could not
recognize them. We involved 80-100 users (in two phases) to help test the
system under everyday circumstances.
The system can provide the final guesses concerning the activities performed
in 2-4 minutes from receiving the data files, and is scalable depending on the
number of users. Based on our back-testing (consisting of a 90-minute
combination of city travel and actions by a small sample of users), the accuracy
of the guesses was 67%.
5. Conclusion
The system presented in this article is suitable for detecting motion in its
most natural form, with devices that are the most widely available. The
conclusion of the research is that it is worthwhile to define several algorithms
that complement each other, in order to improve the accuracy of action
detection.
It is hard to compare the results to the results of other researches in the
literature, because the goal of the research was pretty different from the
previous solutions. In this paper our goal was to recognize the motions in their
immobility walking
stairs
(down)
stairs (up) bus metro tram lift escalator train
immobility 0,00 4,42 4,22 4,49 0,39 0,49 0,60 0,38 0,25 0,22
walking 4,42 0,00 2,46 2,80 4,10 4,02 4,43 4,18 4,28 4,23
stairs
(down)
4,22 2,46 0,00 2,50 3,94 3,82 3,88 3,97 4,07 4,05
stairs (up) 4,49 2,80 2,50 0,00 4,20 4,10 4,40 4,28 4,63 4,32
bus 0,39 4,10 3,94 4,20 0,00 0,27 0,35 0,33 0,24 0,24
metro 0,49 4,02 3,82 4,10 0,27 0,00 0,38 0,32 0,33 0,32
tram 0,60 4,43 3,88 4,40 0,35 0,38 0,00 0,47 0,45 0,44
lift 0,38 4,18 3,97 4,28 0,33 0,32 0,47 0,00 0,30 0,30
escalator 0,25 4,28 4,07 4,63 0,24 0,33 0,45 0,30 0,00 0,15
train 0,22 4,23 4,05 4,32 0,24 0,32 0,44 0,30 0,15 0,00
Unauthenticated
Download Date | 12/19/17 10:39 AM
40 Zs. Sándor, G. Kis
natural way, and not in a laboratory environment. We have used a wide range of
devices, more people, and several device positions. We also tried to recognize
complex activities (such as traveling on tram). The statistical performance of
this system is naturally lower than in a laboratory environment, and also
compared to simpler models aimed at identifying just a few actions. As we
could also see in the literature, accuracy decreases as we allow more and more
flexibility in the system; however, this system, too, becomes more accurate as
actions are aggregated or flexibility is reduced. The closest research found in
the literature is [12], and compared to the 50% accuracy, our 67% is
satisfactory. The results of this research provide satisfactory answers to the
challenges explained at the beginning of the article. A balance has been created
between the need for accuracy and the objective of universality. A system has
been worked out that can be used conveniently by anyone and offers sufficiently
fast response times. Further fine-tuning is underway.
References
[1] Freeman, C., Louca, F., The Emergence of a New Techno-Economic Paradigm: The Age
of Information and Communication Technology (ICT)”, As Time Goes By From the
Industrial Revolutions to the Information Revolution, New York, 2002, pp. 318324.
[2] Naqvi, N. Z., Kumar, A., Chauhan, A., Sahni, K.,“Step Counting Using Smartphone-Based
Accelerometer”, International Journal on Computer Science and Engineering, 4(3), pp.
675681,May. 2012.
[3] Jayalath, S., Abhayasinghe, N., Murray, I.,“A Gyroscope Based Accurate Pedometer
Algorithm”, Presented at International Conference on Indoor Positioning and Indoor
Navigation, Oct. 2013.
[4] Cruz-Silva, N., Mendes-Moreira, J., Menezes, P., Features Selection for Human Activity
Recognition with iPhone Inertial Sensors”, Presented at Portuguese Conference on
Artificial Intelligence, 2013.
[5] Das, S., Green, L., Perez, B., Murphy, M., Detecting User Activities using the
Accelerometer on Android Smartphones, Jul. 2010.
[6] Cerqueira da Silva, J. R., “Smartphone Based Human Activity Prediction”, M.S. thesis,
Faculdade de Engenhariau., Universidade do Porto., 2013.
[7] Siirtola, P., Röning, J., Recognizing Human Activities Userindependently on Smartphones
Based on Accelerometer Data”, International Journal of Interactive Multimedia and
Artificial Intelligence, 1(5), pp. 3845, Jun 2012.
[8] Khan, A. M., Lee, Y.-K., Lee, S. Y., Kim, T.-S., Human Activity Recognition via An
Accelerometer-Enabled-Smartphone Using Kernel Discriminant Analysis”. Presented
atFuture Information Technology, 2010 5th International Conference, May. 2010.
[9] Tomlein, M., Bielik, P., Krátky, P., Mitrík, S., Barla, M., Bieliková, M.,“Advanced
Pedometer for Smartphone-based Activity Tracking”, Presented at Proceedings of the
International Conference on Health Informatics, 2012.
[10] Shin, J., Shin, D., Shin, D., Her, S., Kim, S., Lee, M., “Human Movement Detection
Algorithm Using 3-Axis Accelerometer Sensor Based on Low-Power Management Scheme
for Mobile Health Care System”, Presented at GPC'10 Proceedings of the 5th international
conference on Advances in Grid and Pervasive Computing, May. 2010.
Unauthenticated
Download Date | 12/19/17 10:39 AM
Development of Motion Detection Algorithms 41
[11] Kwapisz, J. R., Weiss, G. M., Moore, S. A., Activity Recognition using Cell Phone
Accelerometers”, ACM SIGKDD Explorations Newsletter, Dec. 2010.
[12] Dernbach, S., Das, B., Krishnan, N. C., Simple and Complex Activity Recognition
Through Smart Phones”, Presented at Intelligent Environments (IE), 2012 8th International
Conference, Jun. 2012.
[13] Yang, C.C., Hsu, Y.L., “Development of a wearable motion detector for telemonitoring and
real-time identification of physical activity”, Telemed. J. E. Health, no. 15, 2009, pp. 6272.
[14] Karantonis, D.M., Narayanan, M.R., Mathie, M., Lovell, N.H., Celler, B.G.,
Implementation of a real-time human movement classifier using a triaxial accelerometer
for ambulatory monitoring”, IEEE. Trans. Inf. Technol. Biomed. no. 10, 2006, pp. 156167.
[15] Ozdemir, A. T. and Barshan, B., Detecting Falls with Wearable Sensors Using Machine
Learning Techniques” Sensors, vol. 14, MDPI, pp. 1069110708, 2014.
Unauthenticated
Download Date | 12/19/17 10:39 AM
Article
Ronald Coase was a founding father of modern law and economics. Yet, Coase distanced himself from the economic analysis of law, which today dominates the law and economics scholarship, and proposed an alternative research program, which is referred to here as “law and the economy”. As market transaction presumes the delineation of rights, which are primarily defined and enforced by law, law obviously anchors the foundation of a market economy. Moreover, changes in the legal system are a main source of institutional change that reassigns rights and redraws constraints under which rights are exercised, thus fundamentally affecting transaction costs and how the economy works. “Law and the economy” recognizes the law as an integral part of the economy and calls for the study of the economic impact of law.
Article
Full-text available
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.
Article
Full-text available
This paper presents a method for counting the number of steps taken by a user, while walking at any variable speed, using smartphone-based accelerometer. For this purpose, the steps are detected based on a relation between frequency of step, which varies inversely with speed of motion, and the magnitude of accelerometer signal. The pattern of the forward acceleration was observed to arrive at the final relation.
Article
Full-text available
Falls are a serious public health problem and possibly life threatening for people in fall risk groups. We develop an automated fall detection system with wearable motion sensor units fitted to the subjects' body at six different positions. Each unit comprises three tri-axial devices (accelerometer, gyroscope, and magnetometer/compass). Fourteen volunteers perform a standardized set of movements including 20 voluntary falls and 16 activities of daily living (ADLs), resulting in a large dataset with 2520 trials. To reduce the computational complexity of training and testing the classifiers, we focus on the raw data for each sensor in a 4 s time window around the point of peak total acceleration of the waist sensor, and then perform feature extraction and reduction. Most earlier studies on fall detection employ rule-based approaches that rely on simple thresholding of the sensor outputs. We successfully distinguish falls from ADLs using six machine learning techniques (classifiers): the k-nearest neighbor (k-NN) classifier, least squares method (LSM), support vector machines (SVM), Bayesian decision making (BDM), dynamic time warping (DTW), and artificial neural networks (ANNs). We compare the performance and the computational complexity of the classifiers and achieve the best results with the k-NN classifier and LSM, with sensitivity, specificity, and accuracy all above 99%. These classifiers also have acceptable computational requirements for training and testing. Our approach would be applicable in real-world scenarios where data records of indeterminate length, containing multiple activities in sequence, are recorded.
Conference Paper
Full-text available
Due to an increased popularity of assistive healthcare technologies activity recognition has become one of the most widely studied problems in technology-driven assistive healthcare domain. Current approaches for smart-phone based activity recognition focus only on simple activities such as locomotion. In this paper, in addition to recognizing simple activities, we investigate the ability to recognize complex activities, such as cooking, cleaning, etc. through a smart phone. Features extracted from the raw inertial sensor data of the smart phone corresponding to the user's activities, are used to train and test supervised machine learning algorithms. The results from the experiments conducted on ten participants indicate that, in isolation, while simple activities can be easily recognized, the performance of the prediction models on complex activities is poor. However, the prediction model is robust enough to recognize simple activities even in the presence of complex activities.
Conference Paper
Full-text available
Accurate step counting is important in pedometer based indoor localization. Existing step detection techniques are not sufficiently accurate, especially at low walking speeds that are commonly observed when navigating unfamiliar environments. This is more critical when vision impaired indoor navigation is considered due to the fact that they have relatively low walking speeds. Almost all existing pedometer techniques use accelerom- eter data to identify steps, which is not very accurate at low walking speeds. This paper describes a gyroscope based pedometer algorithm implemented in a smartphone. The smartphone is placed in the pocket of the trouser, which is a usual carrying position of the mobile phone. The gyroscope sensor data is used for the identification of steps. The algorithm was designed to demand minimal computational resources so that it can be easily implemented in an embedded platform. Raw data from the sensor are filtered using a 6th order Butterworth filter for noise reduction. This is then sent though a zero crossing detector which identifies the steps. A minimum delay between two consecutive zero crossings was used to avoid fluctuations being counted and peak detection was used to validate steps. The algorithm has a calibration mode, in which the absolute minimum swing of data is learnt to set the threshold. This approach demonstrated accuracies above 96% even at slow walking speeds on flat land, above 95% when walking up/down hills and above 90% when going up/down stairs. This has supported the concept that the gyroscope can be used efficiently in step identification for indoor positioning and navigation systems.
Conference Paper
Full-text available
The recognition of human activities through sensors embed-ded in smart-phone devices, such as iPhone, is attracting researchers due to its relevance. The advances of this kind of technology are making possible the widespread and pervasiveness of sensing technology to take advantage of multiple sources of sensing to enrich users experience or to achieve proactive, context-aware applications and services. Human ac-tivity recognition and monitoring involves a continuing analysis of large amounts of data so, any increase or decrease in accuracy results in a wide variation in the number of activities correctly classified and incorrectly classified, so it is very important to increase the rate of correct classifi-cation. We have researched on a vector with 159 different features and on the vector subsets in order to improve the human activities recogni-tion. We extracted features from the Magnitude of the Signal, the raw signal data, the vertical acceleration, the Horizontal acceleration, and the filtered Raw data. In the evaluation process we used the classifiers: Naive Bayes, K-Nearest Neighbor and Random Forest. The features were extracted using the java programming language and the evaluation was done with WEKA. The maximum accuracy was obtained, as expected, with Random Forest using all the 159 features. The best subset found has twelve features: the Pearson correlation between vertical accelera-tion and horizontal acceleration, the Pearson correlation between x and y, the Pearson correlation between x and z, the STD of acceleration z, the STD of digital compass y, the STD of digital compass z, the STD of digital compass x, the mean between axis, the energy of digital compass x, the mean of acceleration x, the mean of acceleration z, the median of acceleration z.
Article
Full-text available
Real-time human activity recognition on a mobile phone is presented in this article. Unlike in most other studies, not only the data were collected using the accelerometers of a smartphone, but also models were implemented to the phone and the whole classification process (preprocessing, feature extraction and classification) was done on the device. The system is trained using phone orientation independent features to recognize five everyday activities: walking, running, cycling, driving a car and sitting/standing while the phone is in the pocket of the subjecttextquoterights trousers. Two classifiers were compared, k nn ( k nearest neighbors) and QDA (quadratic discriminant analysis). The models for real-time activity recognition were trained offline using a data set collected from eight subjects and these offline results were compared to real-time recognition rates, which are obtained by implementing models to mobile activity recognition application which currently supports two operating systems: Symbian^3 and Android. The results show that the presented method is light and, therefore, suitable for be used in real-time recognition. In addition, the recognition rates on the smartphones were encouraging, in fact, the recognition accuracies obtained are approximately as high as offline recognition rates. Also, the results show that the method presented is not an operating system dependent.
Conference Paper
Full-text available
Nowadays many people use smartphones with built-in accelerometers which makes these smartphones capable of recognizing daily activities. However, mobile phones are carried along freely instead of a firm attachment to a body part. Since the output of any body-worn triaxial accelerometer varies for the same physical activity at different positions on a subject's body, the acceleration data thus could vary significantly for the same activity which could result in high within-class variance. Therefore, realization of activity-aware smartphones requires a recognition method that could function independent of phone's position along subjects' bodies. In this study, we present a method to address this problem. The proposed method is validated using five daily physical activities. Activity data is collected from five body positions using a smartphone with a built-in triaxial accelerometer. Features including autoregressive coefficients and signal magnitude area are calculated. Kernel Discriminant Analysis is then employed to extract the significant non-linear discriminating features which maximize the between-class variance and minimize the within-class variance. Final classification is performed by means of artificial neural nets. The average accuracy of about 96% illustrates the effectiveness of the proposed method.