International Journal of Computer Information Systems and Industrial Management Applications.

ISSN 2150-7988 Volume 5 (2013)

(2012) pp. 345-353
c MIR Labs, www.mirlabs.net/ijcisim/index.html

Human Activity Monitoring with Wearable Sensors

and Hybrid Classifiers
Gamze Uslu1 , H.Ibrahim Dursunoglu2 , Ozgur Altun3 and Sebnem Baydere4
Department of Computer Engineering,
Yeditepe University, Istanbul, Turkey
1
guslu, 2 hdursunoglu, 3 oaltun, 4 [email protected]

in the network mentioned. The ultimate goal of an ambient

care system is presenting the assistive technology by meeting the following criteria: Devices in the system should be
embedded in the surroundings or should be wearable. The
constituents of the system should be able to detect the person being serviced and his conditions, the so-called context
awareness principle. The system should also be adjustable
to the personal needs. It should adapt itself depending on
the reactions of the person. It should understand when it is
needed and consequently act as needed without the persons
intrusion, namely principle of being anticipatory.
In the field of ambient care systems and more generally assistive technology, activity monitoring plays a vital role in terms
of taking decisions on when to make the system respond in
what way.If the action performed by a person can be identified, this reveals the information regarding what the person
needs or wants, so that his needs are met by the ambient care
system.The person can be reminded of taking his medication
if he forgets to do so or if the detected action reveals that he
is about to fall, he may be prevented from falling or from a
more severe situation.
There are various technical challenges for the design of activity monitoring systems. Since even the same person does
not perform the same activity in the same way all the time
and some different actions may exhibit similar characteristics, there is a potential deterioration in the recognition accuracy. Noise in the activity signal, namely differentiating
between the noise and the actual signal causes problems as
well. Enhancing an activity monitoring system includes detecting abnormal activities defined in accordance with the
context and providing the appropriate actuation facilities in
response.
Activity monitoring can be achieved in two phases; data collection followed by data classification. Data collection process is carried out through wireless sensors, cameras, PDAs
or other health care monitoring devices[1]. The devices
which do not intrude into the privacy of the person to be monitored can have an advantage over the devices like cameras.
Wireless sensor networks (WSN) can also improve the efficiency of data collection phase. For detection, various classification methods can be used such as least squares[2], knearest neighbor (k-NN)[3], hidden markov, artificial neural
networks (ANN)[4] and support vector machines (SVM)[5]

Abstract: Activity monitoring plays a crucial role in ambient

living environments for assessing changes in the normal behavioral pattern of elderly people. In this paper, we present an action description and detection mechanism for real-time activity
monitoring using wearable sensors and hybrid classifiers. First
a single sensor single classifier model is presented (SSSC) for the
detection of simple and composite actions. Then the model is enhanced with multiple sensors and classifiers for the purpose of
real-time monitoring. The enhanced Multi-Sensor Multi Classifier (MSMC) model uses two wearable TI Chronos watches
with a built-in tri-axial accelerometer for data acquisition and
a composition of naive Bayes, Susan Corner Detector(SCD) and
Hidden Markov(HMM) classifiers for the detection of transitions between defined actions in real-time. A real testbed environment is established to asses the success of real-time monitoring. The test results have revealed that SSSC model is highly
successful in controlled tests when the burden of real-time sampling is ignored whereas MSMC model is fast and accurate for
real time detection of transitions between actions. The proposed
models are tested against the simple actions; walk, sit, stand, lie
as well as walk-while-hands-in-pocket and walk-on-wheelchair.
The unique feature of the selected actions is that the transition between walk, sit and lie are the most likely causes of a
fall event in a home environment for elderly people. The best
achieved detection rates for simple actions range between 92100 % for SSSC model whereas MSMC model is 100 % successful in real-time detection of transitions with a slightly reduced
achievement for individual actions.
Keywords: activity monitoring, hidden markov model, susan corner detection, hybrid classifier, naive bayes, chronos

I. Introduction
Assistive technology provides solutions to people with disabilities and aging population in performing tasks without
being helped by another person. Even if a person is not suffering from disabilities or aging, they still can benefit from
assistive technology tools and services. As a branch of assistive technology, ambient care systems are emerging. To aid
everyday life of people in need, ambient care systems contain
a network of objects used in peoples daily routines. Ambient care systems are capable of sensing the environment
through sensors and reacting to certain conditions reasoned

1
MIR Labs, USA

Uslu, Dursunoglu, Altun and Baydere

346

In this experimental study, an indoor activity monitoring system is designed and implemented to recognize the simple actions performed by a human subject and the transitions between these actions in real time. Two models are analyzed:
single sensor single classifier (SSSC) and multi sensor multi
classifiers( MSMC). In both models, sensor readings from a
tri-axial accelerometer built-in the TI Chronos watch is used
for data acquisition.
In the single sensor-single classifier (SSSC) model, the watch
is worn by the person on the left wrist for the walk action, and
worn to the left thigh for the sit, stand and lie actions. The resulting sensor data obtained in the form of unsigned integers
varying in the range [0,255] are converted to their 2s complement equivalents. The acceleration values in 2s complement form are classified by using naive Bayes classifier into
a simple action. Naive Bayes classifier has training and prediction phases. In the training phase, the training data are exposed to normal distribution to extract unique intervals of average posterior probability. These intervals create the pattern
for the specified action. Patterns for all simple actions are
recorded into a database. In the prediction phase, data sample of a composite action with unknown type is detected by
comparing the differentiated simple actions to the patterns in
the database to produce a posterior probability value. The action of which average posterior probability value is included
in one of the distinct intervals is marked as the corresponding
action.
The multi sensor-multi classifier (MSMC) model includes
a data collection component composed of two acceleration
sensors built in TI Chronos watch worn on wrist and ankle
by the human subject. It also includes a composite data classification subsystem employing naive Bayes classifier first to
differentiate between activities. When naive Bayes classifier
reports that an action performed by a person is ambiguous
ie. may be recognized as more than one action, then the
data classification component uses Hidden Markov Model
(HMM) and Susan corner detector (SCD) in order to find
a single answer to what the action is. Data classification has
training and prediction phases. During the training phase, the
system processes the samples of the actions to be recognized,
generating patterns for those actions. In the prediction phase,
the system processes real time data, splitting the data into
chunks on the fly and inferring what action each chunk belongs to by evaluating the patterns obtained in training phase.
Real-time activities are modeled as simple and composite actions. Simple actions are walk, walk while hands in pocket,
stand, sit, lie and wheelchair driving. Any combination of
these actions are regarded as composite actions. The system is trained with simple actions whereas the real time data
processed during prediction phase are composed of composite actions. The presented work is also featured by its successful differentiation between different kinds of walk actions, namely walk, walk while hands in pocket and walk
on wheelchair.
The content of the following sections is as follows: In Section II related work is reviewed. In Sections III-A and
III-B, single sensor-single classifier and multi sensor-multiclassifier models are given. Section IV reveals the results of
the experiments and elaborates on the future work.

II. Related Work

Human activity monitoring has started to be used in a wide
area. Even though the most common tool used for monitoring activities is camera, it causes computational load as
the number of people being monitored increases. Another
widely used method is PIR sensors. PIR sensors are used
in thermal imaging, radiometry, thermometry and biometry.
Achieving coverage, video surveillance assistance and tracking exploit multiple PIR sensors. A still person can be distinguished from its background with PIR sensors. A video
surveillance system with multi-modal sensor integrity has
been put forward where a tracking system with multiple cameras is united with a wireless sensor network supported with
PIR sensors. Apart from tracking, PIR sensors are beneficial
for detection, differentiation and describing human activity.
The works discussed so far are restricted as a result of the
need for synchronizing the time accurately for sensor nodes
and the related communication cost. There exists a study
which tries to solve these issues by supporting each sensor
node with two PIR sensors to synchronize the data which are
sampled from PIR sensors and communication cost problem
[6].
A method to extract the meaning from the raw sensor data
directly on the sensors is presented by Kay Romer in [7].
A multi-modal sensor system for monitoring human activities is developed by Hung et. al. in [6]. Bao and Intille
developed and evaluated algorithms to detect physical activities and used biaxial accelerometers worn simultaneously on
various parts on the body for data collection in [8]. Zhu and
Sheng propose a human daily activity recognition method by
fusing the data from two wearable inertial sensors attached
on one foot and the waist of the human subject, respectively
in [9]. Lymberopoulos et. al. present an automated methodology for extracting the spatiotemporal activity model of a
person using a wireless sensor network deployed inside a
home in [10]. Bosch et. al. implement and evaluate physical activity monitoring and stimulation using wireless sensor
networks and motion sensors in [11]. Uslu et.al [12] implemented a single sensor single classifier system for monitoring
human activities.
Because data gathered from sensors may not be trustworthy
some researches are carried out targeting sensor data accuracy issue. Hong et. al. address effects of sensor data uncertainty on decision making through information handling
techniques such as Dempster-Shafer theory of evidence and
Equally Weighted Sum operator [13]. Context-aware applications introduce high complexity due to changing context information, changing quality and uncertainty of sensor
data. This complexity necessitates integrating context modelling and reasoning techniques to context aware applications
which also improve maintainability and evolvability along
with reducing complexity. For this reason, Bettini et. al.
compare current context modelling and reasoning techniques
[14]. J. Ye et. al. study situation identification as a way of
coping with uncertainty of sensor data and maps noisy sensor data to patterns useful for applications. They also analyse complexity of situation identification and compare most
common situation identification techniques [15].
Sazonov et. al. present a shoe sensor because distributing
multiple sensors on the body can be too obtrusive. Their

Human Activity Monitoring with Wearable Sensors and Hybrid Classfiers

method can operate without feature extraction to classify

postures and typical activities with respect to heel acceleration and plantar pressure [16]. Chen et. al. employ domain knowledge, ontology and semantic reasoning in activity recognition [17]. People suffering from aging problems
or stroke need continuous physical therapy so activity monitoring systems offer help for physiatrists. Chiang et. al. propose an activity monitoring procedure to assess movements
of patients to see whether they abide what therapy necessitates or not. Their system utilizes WSN based body motion
sensors containing accelerometer and gyroscope along with
fuzzy algorithm to differentiate between static postures and
dynamic motions [18]. Ward et. al. propose performance
metrics for continuous activity recognition [19]. Czabke et.
al. carry out real time activity monitoring with data acquired
from a tri-axial accelerometer and processed on a microcontroller. Their solution does not require a specific positioning
of sensor and they classify actions with no training data [20].
To the best of our knowledge, current studies in the literature
do not particularly focus on capturing transitions between actions on a real-time sequence. Rather they focus on detecting
predefined specific actions.

III. Methodology
We have established an experimental setup to analyze the
proposed SSSC and MSMC models. In both models TI
Chronos watch which contains a 3D accelerometer is used
for the data collection in training and prediction phases. The
watch communicates with the PC through an access point
(AP) over its RF interface operating SimpliciTI protocol
stack. Using multiple Chronos with a single AP brings synchronization problem which causes data loss. For the AP
to receive data from multiple Chronos in MSMC model,
we reprogrammed it. Sensor readings are transmitted from
Chronos to AP after all Chronos establish a connection with
the AP. A master-slave communication protocol is developed between AP and Chronos watches. As a result of this
scheme, while one Chronos transmits to AP, others do not
send any data. In addition, the received data are buffered.
Since Chronos sends twenty acceleration vectors per transmit, the size of the buffer which is the size of a chunk extracted from real time activity data includes twenty vectors.
The classification module of SSSC uses only naive Bayes
classifier whereas MSMC uses naive Bayes classifier as the
initial classifier but upon conflicting detection results, SCD
and HMM are used to resolve the conflict in real-time. The
classifiers naive Bayes and SCD evaluate standard deviation
and mean of training data under normal distribution to generate patterns for actions. The system is designed to monitor
daily activities of a human subject by recording the number
of repeated actions and transitions between actions. Training
of HMM proceeds with calculating the probability of moving to the next state by using the repeated actions and transition counts recorded in the database. In the prediction phase,
naive Bayes classifier processes the real time data to evaluate
the posterior probability value to be compared with the posterior probability of other actions. If the posterior probability
of the real time data corresponds to more than one action,
SCD compares the distance value of the real time data with
other actions. With the condition that SCD is unable to clas-

347

Figure. 1: Formal Representation of Simple and Composite

Actions
sify, HMM assesses transitions, observations and last defined
action. In the following sections, the algorithms used in both
models are explained in detail.
Notations used in Algorithms 1-8 are as follows:
N : Number of simple actions
K : Number of samples taken for each action
S : Number of acceleration vectors in corresponding sample
C : Number of repetitions for an action
card(H) : Number of elements in set H
R : Set of chunks taken from real time data
ACi : Simple action i, with i 1, 2, ..., N
Ttr : Training sample
l : Mean
l : Standard deviation
Al : Array of acceleration vectors along axis l, l {X, Y, Z}
Cz [.] : Test data along Z axis
Tz [.] : Training data along Z axis
A : Array of acceleration vectors
postP rob[.] : Array of posterior probability values
npdf : Normal probability density function
A. Single sensor-single classifier approach
In this approach, continuous activity monitoring is achieved
through detecting successive simple actions in a collected
data samples. The collected data sample from a single
Chronos watch is composed of lines each of which contain
a vector of X,Y and Z axis acceleration values. Simple actions, which can not be split into other actions, are stored in
database and actions contained in collected data are classified
into these actions by utilizing naive Bayes classifier. The defined simple actions are walk, sit, stand and lie, whose content are implied by their names. The collected data sample
holds a composite action since it contains multiple simple
actions. Simple actions which are performed sequentially
produce composite actions. The sample composite actions
experimented are sit-after-lie and walk-after-sit. The formal
definitions of simple and composite actions can be seen as
finite state automata as given in Figure 1. The state S represents a simple action in the automata. At the end of the process, the collected data sample is divided into chunks which
are labeled as one of the actions recorded in the database.
The size of each chunk is determined dynamically and the
size value is used to figure out how long the action related
to that chunk lasts. This is where the innovation of the proposed approach is. Figure 2 illustrates a picture of the test
environment. In this study, all simple and composite action
data samples are obtained in a controlled test environment.
Training phase of the Naive Bayes classifier consists of per-

Uslu, Dursunoglu, Altun and Baydere

348

Figure. 2: Walk and Sit-after-Lie Actions

Figure. 3: Walk-after-Sit Data Sample
forming the same action many times, acquiring distinct intervals for every simple action, finally storing these intervals
in the database. In our experiments, every action is repeated
25 times. The intervals contain average posterior probability
values. Once the training phase is complete, the prediction
process takes place. The algorithm which generates the distinct intervals in training phase is also used in the prediction
phase to generate an average posterior probability value. In
the prediction phase, collected data is processed by extracting a chunk from it in each iteration. n being the current
iteration number, Cn being chunk at iteration n, ci being ith
line in the collected data, iteration number ranging from one
to number of lines in collected data, Cn = [c1 ,c2 ,c3 ,...,cn ]
shows the structure of a chunk extracted from collected data
at any iteration. Chunk at iteration number one is exposed to
the same interval generation scheme in training phase and depending on the resulting average posterior probability value
is in which interval, the chunk is classified into the corresponding action. As long as the following iterations generate the same action classified as the first iteration, the iterative procedure continues, otherwise iteration terminates. The
chunk at the instance of termination is the ultimate structure
representing a simple action classified in collected data and
the iteration number at the instance of termination is the size
of that chunk. After the first chunk is classified, collected
data is truncated so that it starts with the vectors right after
the first chunk and chunk classification continues until there
are no vectors left in collected data.
Algorithm 1 Average posterior probability calculation
S
A Cz [.] Tz [.]
(, 2 ) normalDistribution(A)
postP rob[.] npdf (Tz [.], , 2 )
avg average(postP rob[.])
Average posterior probability of training sample T or a chunk
C from real time sample is calculated as in the Algorithm 1
While naive Bayes classifier can be implemented as assigning the action yielding greatest posterior probability as the
action detected, this work follows the approach that generating unique intervals for every action out of posterior probability values and regards these unique intervals as the differentiating parameter.
There are a number of reasons why normal distribution is
chosen in this work during posterior probability generation
process: First, the normal distribution provides simplicity

Figure. 4: Sit-after-Lie Data Sample

since practically there are many cases where a population

who does not fit normal distribution is successfully processed
under the normal distribution. Second, as the size of the population increases, the probability distribution becomes more
similar to the normal form. Particularly the second point is
a strong reason because the length of the training data can
reach the order of thousands.
Acceleration vs time relationship for the composite action
walk-after-sit are depicted in Figure 3. Here, time is represented virtually by the number of the acceleration vectors
collected during the action. Higher number of vectors resembles an action of longer duration. With the transmission
frequency of 33 packets per second, approximately 2.5 s corresponds to 45 vectors. During this approximate calculation,
lost packets caused by Chronos are ignored. Figure 3 illustrates the collected data sample for the walk-after-sit action.
In these plots, the action occurs as the combination of simple
actions sit, stand and walk respectively. Since the aim is detecting walk after sit rather than the sequence of sit, stand and
walk, during the tests, the section of the signal showing the
stand action is ignored by filtering that segment, regarding
stand as the transition. The data is obtained by appending
walk samples to the end of sit samples. The training data
related to the simple walk action are collected by wearing
sensor on the left wrist whereas the composite action is performed with the sensor on the left thigh. The transition signal
samples for the sit-after-lie action are also illustrated in Figure 4.

Human Activity Monitoring with Wearable Sensors and Hybrid Classfiers

349

acceleration vs time
60
X
Y
Z

40

20

acceleration

20

40

60

80

100

(a) Wheelchair Test

(b) Stand Test

120

140

50

100

150

200

250

time

Figure. 5: Sensor placement on human body

(a) Walking test data sample for arm sensor

1) Training

2) Prediction
A collected data sample of an action whose type is unknown
is compared to all of the action patterns in the database, producing a posterior probability value. In posterior probability
calculation, normal probability density function values found
for the Z axis are used. To evaluate a normal probability density function value, the acceleration data obtained from the
collected data sample, the mean and standard deviation values related to the specified axis are used.
B. Multi sensor-multi classifier approach
MSMC model focuses on the detection of transitions. Detecting transition between activities can not always be done
with SSSC model. In the cases where SSSC model is insufficient to generate a unique class for an action, using multiple
sensors help generating such a unique class by enlarging the
feature space. Thus the use of multiple sensors decreases the
probability of an action overlapping with the unique features
of more than one action. Also, hybrid classifiers have the
potential to improve the detection success further.
Algorithm 2 Main classification module
for i <= N do
vectorCount 0
for j <= K do
for m <= S do
vectorCount vectorCount + 1
A[vectorCount] Ttr [m]
end for
end for
end for
Actions are classified according to the location of Chronos
watch yielding the signal of that action. One Chronos is worn
on right ankle whereas the other one is worn on right wrist
as in Figure 5(a) and 5(b). Plots of walking and wheelchair
driving are shown in Figure 6(a), Figure 6(b), Figure 7(a) and

X
Y
Z
100

50

acceleration

Training data are exposed to normal distribution to extract

mean and standard deviation for the Z axis. These values are
used to calculate average posterior probability which form
the pattern of an action. A pattern is created for every action
and all patterns are inserted to a database to be used in the
prediction phase later.

acceleration vs time
150

50

100

150

20

40

60

80

100
time

120

140

160

180

200

(b) Walking test data sample for foot sensor

Figure. 6: Walking real time test data samples

Figure 7(b). The sensor on right ankle forms the foot oriented
detection subsystem and the sensor on the right wrist forms
hand oriented detection subsystem. System distinguishes
the Chronos watches according to first bit of dataset. Signals emitted from both Chronos are combined after separate
recognition.
The main classification module is shown in Algorithm 2.
This algorithm computes the mean and standard deviation
of the training data along X, Y and Z axes.
In the prediction phase of naive Bayes classifier as shown
in Algorithm 4, a chunk from real time data is compared to
the unique intervals generated for each action. If a single
class is obtained for that chunk, detection becomes complete,
otherwise classification proceeds with HMM or SCD.
Training phase of naive Bayes classifier as given in Algorithm 3 calculates the posterior probability value, namely
pattern Pj for each sample taken for every action through
mean and standard deviation of the acceleration data
along X and Z axes since these are the dominant axes for the
actions within the scope of this work. Range[i,j] indicates

Algorithm 3 Training phase of naive Bayes classifier

for i = 1 N do
for j = 1 K do
Pj f (i(X,Z) , i(X,Z) );
Range[i, j] Pj ;
end for
end for

Uslu, Dursunoglu, Altun and Baydere

350

Algorithm 4 Prediction phase of naive Bayes classifier

for all jR do
countN onM atching 0
countM atching 0
for i = 1 N do
Pj f (i(X,Z) , i(X,Z) )
if ( thenPj Range[i])
countM atching countM atching + 1
else
countN onM atching count + 1
end if
end for
if countN onM atching = N then
call HMM scheme
else
if countM atching > 1 then
call SCD scheme.
else
Classify segment j as action i
end if
end if
end for

acceleration vs time
60
X
Y
Z

40

20

acceleration

20

40

60

80

100

120

140

20

40

60

80

100
time

120

140

160

180

200

(a) Wheel chair test data sample for arm sensor

acceleration vs time
100
X
Y
Z

50

acceleration

50

100

the range of posterior probability values which belong to the

action i, j indicating the sample index related to the action i.
SCD algorithm maps data to a nucleus whose center is equal
to Y and radius is equal to Y . Algorithm 5 and 6 explain
how SCD scheme fulfils classification in training and prediction phases.
The chunks from real time data which can not be defined
with naive Bayes classifier and SCD are exposed to HMM
scheme. HMM considers two events last action and next
action. Last action is the last detected action whereas next
action is the current action to be detected. If the last action is
not defined, as the last action HMM assigns the action with
greatest observation probability, in our case sit whose observation probability is shown in Table 1.
Having an action as the last action is important since HMM
needs to find the transition probability from last action to next
action. After assigning last action, HMM estimates the next
action using (1).
L(n|l, o) = P (n|l).P (n|o)

(1)

Interpretation of (1) is as follows: L(n|l, o) indicates probability that the next action is n, given that l is the last action
and o is the observation probability of last action. P (n|l)
shows the probability of obtaining n as the last action given
that last action is l. Finally, P (n|o) is the probability of obtaining n as the last action given that o is the observation
probability of last action. HMM checks transition probabilities from last action to every action. Transition probabilities
between actions are depicted in Table 2 where transition column indicates actions between which the transition occurs

Table 1: Observation probabilities of simple actions

Action Name
Walk
Sit
Stand

Observation Probability
0.35
0.45
0.2

150

20

40

60

80

100
time

120

140

160

180

200

(b) Wheel chair test data sample for foot sensor

Figure. 7: Wheel chair real time test data samples

such that Walk-Sit means the transition from walk to sit. In

the case of potential next actions that generate equal transition probabilities, the action which yields greatest observation probability is selected to be the next action. HMM training and prediction phases are presented in Algorithm 7 and 8.
In these algorithms, transProb[.] and obsProb[.] show transition probability matrix and observation probability vector for
simple actions. Also, transition probability of moving from
action i to action j is shown as transProb(i,j) and maximum of
the transition probabilities from action i to all other actions
is designated as max(transProb(i),Y), Y being the set of actions generating the maximum transition probability. Finally,
maximum of the observation probabilities of the actions is
indicated by max(obsProb,Y), Y being the action having the
maximum observation probability.

Algorithm 5 Training phase of SCD

for i = 1 N do
vectorCount 0
for j = 1 K do
for m = 1 S do
distance=([X,Y,Z]-[0,center,0])/radius
vectorCount++
Range[i,vectorCount]=distance
end for
end for
end for

Human Activity Monitoring with Wearable Sensors and Hybrid Classfiers

Algorithm 6 Prediction phase of SCD

actionF ound 0
for i = 1 N do
distance=([X,Y,Z]-[0,center,0])/radius
if distance
/ Range[i] then
SCD fails to classify the segment
else
actionF ound 1
Mark action as the action corresponding to i
Terminate SCD
end if
end for
if actionF ound = 0 then
SCD fails to classify this segment
Call HMM scheme.
end if

351

Algorithm 8 Prediction phase of HMM

if lastAction
/ def ined then
lastAction sit
end if
max(transProb(i),Y)
if card(Y ) > 1 then
max(obsProb,Y)
Segment is classified as Y
else
a Y
Segment is classified as a
end if
chunk. The tests are repeated several times. On average following success rates are achieved: walk 92%, sit 100%, lie
88% and stand 96% as tabulated in Table 3. Based on these
simple actions, various numbers of tests are performed for
the detection of walk-after-sit and sit-after-lie composite actions. The model classified the actions successfully when
the transition signals that normally occur between actions in
real-life samples are ignored.

Algorithm 7 Training phase of HMM

transP rob[.] 0
obsP rob[.] 0
for i = 1 N do
for t = 1 C do
for j = 1 N do
Table 3: Detection success rates for simple actions (SSSC)
action=ACi ACj
Action Name
Detection Success Rate
(transAction1 ,transAction2 )=nBayes(action)
Walk
92%
transProb(transAction1,transAction2 )++
Sit
100%
obsProb(transAction2)++
Lie
88%
end for
Stand
96%
end for
end for
In multi sensor-multi classifier approach, a hybrid classifier
is used. naive Bayes, HMM and SCD are used to detect transitions between activities in real time. In the training stage of
IV. Results and Conclusion
Naive Bayes classifier, data are processed to calculate mean
and standard deviation under the normal probability distriThis experimental study covers the design and implementa- bution to extract a posterior probability. In addition to this,
tion of a real time indoor human activity monitoring system every action and transition between the actions is observed
by addressing the two phases of activity monitoring, namely to calculate transition probability and observation probability
data acquisition and classification. Two models are used under joint probability distribution in the training of HMM.
in the experiments; single sensor-single classifier and multi In the prediction phase, the posterior probability of the real
sensors-multi classifiers models.
time data is calculated. If it overlaps with any other action,
the data are processed by SCD to find distance to center of
circle. If the distance is mapped to any action, it is marked as
Table 2: Transition probabilities of simple actions
Action Name
Transition Probability
detected action. Otherwise, HMM classifier finds the next acWalk-Walk
0.8
tion using transition and observation probabilities, obtained
Walk-Sit
0.05
in
training stage. The tests are repeated for several times and
Walk-Stand
0.15
it
has
been observed that the model is fast and 100% accurate
Sit-Walk
0.2
Sit-Sit
0.6
in detecting the transition signals. However, because of the
Sit-Stand
0.2
real time processing delay a minor reduction in the detection
Stand-Walk
0.3
of individual actions are observed.
Stand-Sit
0.2
Stand-Stand

0.5

Table 4: Detection success rates for simple actions(MSMC)

In single sensor-single classifier approach, Naive Bayes classifier is implemented for the data classification subsystem.
Unique intervals of average posterior probability of the training data in 2s complement form are calculated with the
normal distribution in the training phase. In the prediction
phase, the real time sample is partitioned into chunks at the
points where chunks show the simple actions and the chunk
size shows the representative duration of classified action by
means of calculating average posterior probability for each

Action Name
Walk
Walk while hands in pocket
Sit
Stand
Wheelchair driving

Detection Success Rate

94%
96%
94%
94%
98%

Detection success rates achieved with MSMC model are 94%

for walk, 96% for walking while hands in pocket, 94% for
sit, 94% for stand, and 98% for wheelchair driving, as given

352

in Table 4. The results reveal that the proposed filtering

mechanism can successfully distinguish actions. As a future
work, the model will be integrated into an abnormal activity
detection scenario, such as fall and bump.

References
[1] Y. Lin, E. Becker, K. Park, Z. Le, and F. Makedon,
Decision making in assistive environments using multimodal observations, in PETRA 09 Proceedings of the
2nd International Conference on PErvsive Technologies Related to Assistive Environments. ACM, 2009,
pp. 18.
[2] D. W. Marquardt, An Algorithm for Least-Squares
Estimation of Nonlinear Parameters, SIAM Journal
on Applied Mathematics, vol. 11, no. 2, pp. 431441,
1963.
[3] P. Cunningham and S. J. Delany, k-Nearest Neighbour Classifiers, Multiple Classifier Systems, pp. 117
(UCDCSI20074), University College Dublin, 2007.
[4] J. Hopfield, Artificial neural networks, Circuits and
Devices Magazine, IEEE, vol. 4, no. 5, pp. 3 10, sep
1988.
[5] S. R. Gunn, Support Vector Machines for Classification and Regression, The Analyst, vol. 135, no. 2, pp.
230267, Citeseer, 1998.
[6] Hung, Tahir, Farrell, McLoone, and McCarthy, Wireless Sensor Networks for Activity Monitoring using
Multi-sensor Multi-modal Node Architecture, Proceedings of Signals and Systems Conference (ISSC
2009), IET Irish, pp. 16, 2009.
[7] K. Romer, Discovery of Frequent Distributed Event
Patterns in Sensor Networks, in European Conference
on Wireless Sensor Networks EWSN, vol. 67322, no.
5005. Springer-Verlag, 2008, pp. 106124.
[8] L. Bao and S. S. Intille, Activity Recognition from
User-Annotated Acceleration Data, Proceedings of
Pervasive Computing, Second International Conference, PERVASIVE 2004, vol. 3001, pp. 117, Springer,
2004.
[9] C. Zhu and W. Sheng, Human Daily Activity Recognition in Robot-assisted Living Using Multi-sensor Fusion, Proceedings of Robotics and Automation, 2009,
ICRA 09, IEEE International Conference, pp. 2154
2159, 2009.
[10] D. Lymberopoulos, A. Bamis, and A. Savvides, Extracting spatiotemporal human activity patterns in assisted living using a home sensor network, Proceedings of the 1st ACM international conference on PErvasive Technologies Related to Assistive Environments
PETRA 08, p. 1, 2008.
[11] S. Bosch, M. Marin-Perianu, R. Marin-Perianu,
P. Havinga, and H. Hermens, Keep on moving! activity monitoring and stimulation using wireless sensor

Uslu, Dursunoglu, Altun and Baydere

networks, Proceedings of EuroSSC09, the 4th European conference on Smart sensing and context, pp. 11
23, ACM, 2009.
[12] G. Uslu, O. Altun, and S. Baydere, A Bayesian Approach for Indoor Human Activity Monitoring, Proceedings of IEEE Int. Conference on Hybrid Intelligence Systems, HIS 2011, pp. 324327, 2011.
[13] X. Hong, C. Nugent, M. Mulvenna, S. McClean,
B. Scotney, and S. Devlin, Evidential fusion of sensor
data for activity recognition in smart homes, Pervasive
and Mobile Computing, vol. 5, no. 3, pp. 236 252,
Elsevier Science Publishers B. V., 2009.
[14] C. Bettini, O. Brdiczka, K. Henricksen, J. Indulska,
D. Nicklas, A. Ranganathan, and D. Riboni, A survey
of context modelling and reasoning techniques, Pervasive Mob. Comput., vol. 6, no. 2, pp. 161180, Elsevier
Science Publishers B. V., Apr. 2010.
[15] J. Ye, S. Dobson, and S. McKeever, Situation identification techniques in pervasive computing: A review,
Pervasive and Mobile Computing, vol. 8, no. 1, pp. 36
66, Elsevier Science Publishers B. V., 2012.
[16] E. Sazonov, G. Fulk, J. Hill, Y. Schutz, and R. Browning, Monitoring of posture allocations and activities by
a shoe-based wearable sensor, Biomedical Engineering, IEEE Transactions on, vol. 58, no. 4, pp. 983 990,
april 2011.
[17] L. Chen, C. Nugent, and H. Wang, A knowledgedriven approach to activity recognition in smart
homes, Knowledge and Data Engineering, IEEE
Transactions on, vol. 24, no. 6, pp. 961 974, june
2012.
[18] S.-Y. Chiang, Y.-C. Kan, Y.-C. Tu, and H.-C. Lin, Activity recognition by fuzzy logic system in wireless sensor network for physical therapy, in Intelligent Decision Technologies, ser. Smart Innovation, Systems
and Technologies, J. Watada, T. Watanabe, G. PhillipsWren, R. J. Howlett, and L. C. Jain, Eds., Springer
Berlin Heidelberg, 2012, vol. 16, pp. 191200.
[19] J. A. Ward, P. Lukowicz, and H. W. Gellersen, Performance metrics for activity recognition, ACM Trans.
Intell. Syst. Technol., vol. 2, no. 1, pp. 6:16:23, Jan.
2011.
[20] A. Czabke, S. Marsch, and T. Lueth, Accelerometer
based real-time activity analysis on a microcontroller,
in Pervasive Computing Technologies for Healthcare
(PervasiveHealth), 2011 5th IEEE International Conference on, may 2011, pp. 40 46.

Human Activity Monitoring with Wearable Sensors and Hybrid Classfiers

Author Biographies
Gamze Uslu received her BSc degree in
computer engineering from Yeditepe University, Istanbul, Turkey in 2011. She is
a research assistant and currently pursuing
the MSc degree in computer engineering at
Yeditepe University. Her research interests
include activity monitoring, machine learning and body sensor networks.
Halil Ibrahim Dursunoglu received the
B.Sc. degree in Computer Engineering from
Yeditepe University, Istanbul, Turkey, in
2012. He works as manager in an international company. His research interests include wireless sensor networks and activity
monitoring.
Ozgur Altun received the B.Sc. degree in
Computer Engineering from Kocaeli University, Turkey, in 2010. He is currently pursuing the M.Sc. degree in Satellite Communication and Remote Sensing at Istanbul
Technical University. He works as a research assistant at Yeditepe University. His research interests
include wireless sensor networks, activity monitoring, wireless and satellite communication.

353

Sebnem Baydere is a full Professor and the

Chair of the Department of Computer Engineering at Yeditepe University, Istanbul. She
received her BSc and MSc degrees in Computer Engineering from Middle East Technical University (METU), Ankara, in 1984
and 1987 respectively. She received her PhD degree in Computer Science from University College London (UCL), UK,
in 1990. Her current research interests are in the area of wireless sensor and ad hoc networks, wireless multimedia networks, network interoperability, context aware systems and
distributed systems. Dr Baydere coordinated several national
and international research projects and published papers in
the area of wireless sensor networks and distributed systems.
She served as technical committee member for several conferences and peer-reviewed journal articles on Sensor Networks. She also served numerous times as evaluator for European and National research projects.