scispace - formally typeset
Search or ask a question
Proceedings ArticleDOI

Accelerometer Placement for Posture Recognition and Fall Detection

25 Jul 2011-pp 47-54
TL;DR: Chest and waist accelerometers proved best at both tasks, with the chest accelerometer having a slight advantage in posture recognition.
Abstract: This paper presents an approach to fall detection with accelerometers that exploits posture recognition to identify postures that may be the result of a fall. Posture recognition as a standalone task was also studied. Nine placements of up to four sensors were considered: on the waist, chest, thigh and ankle. The results are compared to the results of a system using ultra wide band location sensors on a scenario consisting of events difficult to recognize as falls or non-falls. Three accelerometers proved sufficient to correctly recognize all the events except one(a slow fall). The location-based system was comparable to two accelerometers, except that it was able to recognize the slow fall because it resulted in lying outside the bed, whose location was known to the system. One accelerometer was able to recognize only the most clear-cut fall. Two accelerometers achieved over 90% accuracy of posture recognition, which was better than the location-based system. Chest and waist accelerometers proved best at both tasks, with the chest accelerometer having a slight advantage in posture recognition.

Summary (2 min read)

Introduction

  • With accelerometers that exploits posture recognition to identify postures that may be the result of a fall.
  • Improving the quality of life of Europe’s increasing elderly population is one of the most pressing challenges facing their society.
  • The authors also investigated the performance of posture recognition and fall detection with different numbers of accelerometers (1 to 4) and different placements on the body (chest, waist, ankle and thigh).
  • Eventually a comparison with a system based on location sensors was made.
  • They used two 3-axis accelerometers and gyroscopes worn on the chest and thigh; by applying thresholds to accelerations, angular velocities and angles, they detected a potential fall and the activity/posture after the fall, resulting in 90.1% accurate fall detection.

B. System Architecture

  • The posture recognition and fall detection process described in this paper is divided in several phases .
  • The data mining posture recognition module classifies the acceleration data samples into 7 predefined postures.
  • The first algorithm in this module detects high accelerations.
  • The number and placements of the accelerometers on the body can affect the recognition of particular postures.
  • Therefore, successive activities (e.g. standing -> going down -> lying) may have some samples mislabeled at the beginning and at the end of each posture/activity.

A. Body Posture Recognition Module

  • The real-time posture recognition module uses a classification model to classify every data sample that is received from the accelerometers.
  • Additional 18 attributes are extracted for each accelerometer: Length of the acceleration vector (1 attribute) Orientation angles for each axis – x, y and z (3 attributes) Statistical attributes for each axis and for the length of the acceleration vector o Mean Value (4 attributes) o Root Mean Square (4 attributes) o Standard Deviation (4 attributes).
  • This attribute improves the classification for static postures that have different sensor angle orientations.
  • The first attribute from this group is the Mean value of the data in the window.
  • Because the authors recorded 11 people, the model was trained on 10 people and tested on the remaining person.

B. Fall Detection Module

  • The second module is focused on the detection of fall events.
  • In this module two acceleration-based algorithms are used.
  • The threshold was chosen empirically for each of the accelerometers individually.
  • To solve this issue the authors used a second algorithm, which takes into account the recognized posture after a potential fall event.
  • The fall had occurred if: Acceleration has exceeded the threshold as described in the first algorithm AND the person is lying for more than 10 seconds; OR The person sits on the ground more than 10 seconds.

A. Test Scenario

  • The test scenario is around 15 minutes long and includes all the target body postures.
  • They were recorded in single recordings interspersed with short periods of walking.
  • As shown in the section on related work, accelerometers can accurately detect typical falls, so the authors included only one such fall (event number 2) to demonstrate that the system can recognize it accurately.
  • Furthermore, the authors included two events (5 and 7) that involve high acceleration and could thus be misclassified as falls by accelerometers.
  • The last two events (1 and 3) are perfectly normal and were included to verify that all the methods work correctly and do not recognize them as falls.

B. Results

  • The results are shown for each of the modules in separate and different ways.
  • Three accelerometers offer 7–9 percentage points better accuracy.
  • Looking at the average accuracies and other false alarms, one can make the following conclusions:.
  • With one accelerometer placed on the chest or waist, the classification model was able to distinguish between two groups of postures ({lying and on all fours} and {standing, sitting and sitting on the ground}).
  • The accelerometers fall detection performance can be improved by the information about the location where a potential fall is taking place.

Did you find this useful? Give us your feedback

Content maybe subject to copyright    Report

Accelerometer Placement for Posture Recognition
and Fall Detection
Hristijan Gjoreski, Mitja Luštrek, Matjaž Gams
Department of Intelligent Systems
Jožef Stefan Institute
Jamova cesta 39, 1000 Ljubljana, Slovenia
E-mail: {hristijan.gjoreski, mitja.lustrek, matjaz.gams}@ijs.si
AbstractThis paper presents an approach to fall detection
with accelerometers that exploits posture recognition to identify
postures that may be the result of a fall. Posture recognition as a
standalone task was also studied. Nine placements of up to four
sensors were considered: on the waist, chest, thigh and ankle. The
results are compared to the results of a system using ultra-
wideband location sensors on a scenario consisting of events
difficult to recognize as falls or non-falls. Three accelerometers
proved sufficient to correctly recognize all the events except one
(a slow fall). The location-based system was comparable to two
accelerometers, except that it was able to recognize the slow fall
because it resulted in lying outside the bed, whose location was
known to the system. One accelerometer was able to recognize
only the most clear-cut fall. Two accelerometers achieved over
90% accuracy of posture recognition, which was better than the
location-based system. Chest and waist accelerometers proved
best at both tasks, with the chest accelerometer having a slight
advantage in posture recognition.
Keywords Ambient intelligence, fall detection, posture
recognition, activity recognition, accelerometers, accelerometer
placement, classification.
I. INTRODUCTION
There are many studies describing new ways of improving
the life of elderly. Improving the quality of life of Europe’s
increasing elderly population is one of the most pressing
challenges facing our society. Nearly 14% of the EU
population is over 65 and this figure is expected to double by
2050 [8]. By then Europe will have 80 million elderly citizens
who should continue to play an active role in our society,
despite limitations which the ageing process often brings. To
cope with this situation, intelligent systems and techniques are
being developed, which can give these people self-confidence
to actively and independently live their life despite their age
limitations.
The detection of falls is important task in ambient assisted
living. Furthermore, activity/posture recognition is an essential
component in such systems. In this paper we present our work
on the detection of falls and alarming situations using wearable
accelerometers. Fall detection is improved by body posture
recognition. In posture recognition we were focused on seven
basic body postures. We also investigated the performance of
posture recognition and fall detection with different numbers of
accelerometers (1 to 4) and different placements on the body
(chest, waist, ankle and thigh). This way we showed the
tradeoff between the intrusiveness of the system and the
achieved accuracy. The final system should be as non-intrusive
as possible (fewer wearable sensors), but still accurate enough
to detect each fall. Eventually a comparison with a system
based on location sensors was made. This system was
developed in a European FP7 project - Confidence [11].
II. RELATED WORK
Lots of studies in ambient intelligence, particularly in
ambient assisted living are focused on detecting alarming
situations using different kinds of technical equipment. Many
of the studies on human body posture analysis and fall
detection use image processing, location sensors or
accelerometers. However, each of these approaches has certain
problems.
The image processing approach [6] has several operational
complexities. One of them is the process of installation of the
camera in each room we want the system to work. Another
issue is limitation of functioning only in indoor environment.
Also common disadvantages are the low image resolution and
target occlusion. And probably the biggest issue in this
approach is the user privacy. The user has to accept the fact
that a camera will record him/her.
Some researchers have tried to analyze the body posture
and fall detection using location sensors [3, 9, 10, 11]. The
problem with this approach is the variation of the sensors
precision (not constant rate of distance error) and price of the
sensors. For reasonable results high precision is required
(Ubisense 2011). High precision means high cost of sensors.
This price is many times higher than the price of
accelerometers. In addition, accelerometers are far more
commercially available and nowadays they can be found
almost in any smart phone. Similarly to the image processing
approach, these location systems are limited to indoor
environment and require installation in each room where we
want the system to work. The advantage of location sensors
compared to accelerometers is knowing the location of the
person. That can be seen in our comparison with the location-
based system [11] presented in the fall detection results
section.

Another way of formulating posture recognition and fall
detection task is by using accelerometers.
The most common approach for posture recognition is the
data mining [1, 2]. In a narrow sense it can be interpreted as a
pattern recognition problem [1]. Therefore good attributes that
explain the body posture are essential in this approach.
Usually the results are presented in detection of the whole
process of the activity/posture (e.g. lying for 20 seconds as
one sample). In our work we tried to predict each data sample
and the results are presented that way.
Also there are studies that use manually created algorithms
to formulate the accelerometer based posture recognition task
[13]. With this approach the achieved accuracies are very
good (around 99%), but it requires using multiple
accelerometers (6).
Most studies on fall detection use accelerometers (which
measure linear acceleration) and gyroscopes (which measure
angular velocity). Some researchers used machine learning
instead of threshold-based algorithms [14, 15]. In these 2
approaches they used a triaxial accelerometer worn on the
waist. Using SVM machine learning algorithm on various
features derived from accelerations, they detected falls with
96.7 % and 100 % accuracy, respectively.
Typical approach is detection of falls by applying
thresholds to accelerations, velocities and angles [7]. In [12]
they used a 3-axis accelerometer worn on the chest; by
applying a simple threshold to the acceleration, they detected
falls with 98.9% accuracy. In [4] they used a 3-axis
accelerometer worn on the waist; by applying thresholds to the
acceleration, they detected a potential fall and the
activity/posture after the fall, resulting in 100% accurate fall
detection.
Of particular interest to us is the work in [5]. They used
two 3-axis accelerometers and gyroscopes worn on the chest
and thigh; by applying thresholds to accelerations, angular
velocities and angles, they detected a potential fall and the
activity/posture after the fall, resulting in 90.1% accurate fall
detection. The lower accuracy compared to the previous work
is most likely due to the more difficult test data: their method
sometimes failed on lying down quickly and on two atypical
fall types. Exactly such situations are tested in this paper. Fall
detection using accelerometers may appear straightforward,
but detecting all types of falls is challenging. Most of the
researches show only accuracy in detecting fast falls. Because
of that, having a good testing scenario which includes all kinds
of falls is one of the key steps in fall detection procedure. In
our approach we were interested in different alarming
situations that were more difficult to detect only acceleration
signal: falling slowly (e.g. losing consciousness) and sliding
from chair (quickly and slowly). In addition, we tested sitting
down quickly. All these events are included in the test
scenario.
III. PROBLEM DESCRIPTION
In this section we describe the hardware components used
in this research, and the architecture of our system for posture
recognition and fall detection.
A. Hardware
The sensors used in this research were four 3-axis
accelerometers. A 3-axis accelerometer is a sensor that returns
a real-valued estimate of the acceleration along the x, y and z
axes from which velocity and orientation angle can also be
estimated. Accelerometers measure the acceleration and output
the projections of the acceleration vector represented in a three
dimensional coordinate system. Each accelerometer has its own
coordinate system and gives the relative vector projections.
Because of the Earth’s gravity, all objects experience a
gravitational pull towards the Earth’s center. The acceleration
unit of the pull is referred to as g or g force. Consequently all
objects are subject to 1 g acceleration. When the accelerometer
is at rest, only Earth’s gravity is measured. Accelerometers can
be used as motion detectors as well as for body-posture
recognition and fall detection. Acceleration sensors used in this
research have data sampling frequency of 6 Hz. This is not a
high frequency, but it gives enough information for our final
goal and makes the system more compact and portable even on
devices with low memory and low processing power. It is also
equal to the sampling frequency of the location sensors we
used for comparison.
B. System Architecture
The posture recognition and fall detection process described
in this paper is divided in several phases (Figure 1). It begins
with the real world in which wearable accelerometers are
affixed to a person’s body. The next phase is the sensory part.
Accelerometers sample the signal and send the raw data to the
software attribute extraction module. The software analyzes the
raw data and extracts new attributes. To solve the problem of
posture recognition and detecting fall situations two software
modules were developed: data mining body posture recognition
and fall detection. Each of these modules gives an output: the
real time posture of the body and a fall or non-fall event.
Figure 1: System Architecture
The data mining posture recognition module classifies the
acceleration data samples into 7 predefined postures. Five of
the postures were chosen to be common and general: standing,
sitting, lying, standing up and going down (which includes
sitting down, lying down or falling). Two of the postures are
more specific and are related to fall detection: on all fours and
sitting on the ground. The process of walking is included into
posture standing, therefore standing can be a static or a
dynamic posture.
As we already mentioned, each accelerometer experiences
the gravitational acceleration. This is an important reference
for calculating sensor orientation angle and distinguishing
different body postures. The orientation angle is extracted as
the angle between the acceleration vector (g if the posture is
static) and each of the axis unit vectors (x, y and z). Using

these 3 angles we have a unique sensor space orientation. With
the extraction of the orientation angle attribute, the data
mining algorithm can easily distinguish static postures that
have different sensor orientation angles. For instance the chest
accelerometer is perpendicular to the ground for standing and
sitting postures, but parallel to the ground for lying and on all
fours postures. Because of such similarities and differences in
the orientation of the sensors the body postures can be
recognized. Also a special movement attribute is extracted for
distinguishing dynamic from static postures. This attribute
captures the changes in the acceleration vector; the greater the
changes, the fiercer the motion.
The fall detection module is focused on the detection of
fall events. The first algorithm in this module detects high
accelerations. It is a threshold-based algorithm which has
some improvements for reducing false alarms. The second
algorithm uses the information from posture-recognition
module as input. It analyzes the recognized postures and
decides if a fall occurred.
The number and placements of the accelerometers on the
body can affect the recognition of particular postures. Placing
an accelerometer on the thigh can help distinguishing sitting
(accelerometer is parallel to the ground) and standing
(accelerometer is perpendicular to the ground), but
distinguishing sitting and lying is a problem. On the other
hand, an accelerometer on the chest can distinguish sitting and
lying, but has problems with standing and sitting. By
combining these two accelerometers the algorithm was able to
distinguish each of the discussed postures. Because of
situations like this, we decided to compare the results using
different numbers of accelerometers and different body
placements. The idea is to use as few sensors as possible to
maximize the user’s comfort, but to use enough of them to
achieve satisfactory performance. It is very important that the
accelerometers are firmly fixed, because the orientation of the
sensor is the most important feature in posture recognition. The
body placements of the accelerometers were chosen to be:
chest, waist, right thigh and the right ankle. The chest and waist
positions are common in fall detection, so we wished to
compare them.
IV. METHODS
In this section we describe the methods and algorithms used
in each of the modules in our system. Because we use data
mining classification techniques for body posture recognition,
the data collecting is one of the main steps. We use supervised
classification techniques, therefore the data had to be labeled
with the appropriate posture. As the raw data sample was
received from the sensors, firstly it was labeled with the
appropriate class value (posture) and afterwards saved in a
database together with the correct posture. This was done
manually while recording. We should note that because the
labeling was done online, small delays in labels could happen.
Therefore, successive activities (e.g. standing -> going down ->
lying) may have some samples mislabeled at the beginning and
at the end of each posture/activity. This can affect the
recognition accuracy especially for short postures/activities
(going down or standing up) that do not have lots of samples.
In addition, the border between two successive postures is to
some degree subject to interpretation.
A. Body Posture Recognition Module
The real-time posture recognition module uses a
classification model to classify (recognize) every data sample
that is received from the accelerometers. Therefore good
attributes that describe the person’s posture are essential for
successful classification. The first set of attributes are the raw
acceleration vector projections (3 attributes for each
accelerometer) a
x
, a
y
and a
z
. Additional 18 attributes are
extracted for each accelerometer:
Length of the acceleration vector (1 attribute)
Orientation angles for each axis x, y and z (3 attributes)
Statistical attributes for each axis and for the length of the
acceleration vector
o Mean Value (4 attributes)
o Root Mean Square (4 attributes)
o Standard Deviation (4 attributes)
Movement detection attributes (2 attributes)
The first derived attribute is the length (module) of the
acceleration vector. It is a simple but very useful attribute,
which is also used further in the process of extraction of new
attributes. Its definition is:
.
2
z
2
y
2
x
aaalength
(1)
During static postures this attribute is constant with the value
equal to the Earth’s gravity (M = 1g). Otherwise in dynamic
activities the acceleration vector is changing the direction and
its module.
The most important characteristics for static body posture
recognition are the orientation angles of the accelerometer.
The orientation angles are calculated as the angles between the
actual acceleration (Earth’s gravity for static postures) and
each of the axes (x, y and z). For instance, the angle φ
x
between the acceleration vector and the x axis (perpendicular
to the ground) is computed as follows:
.cos
2
z
2
y
2
x
aaa
a
x
x
(2)
This attribute improves the classification for static postures
that have different sensor angle orientations. Since each
person has his/her characteristic posture and each
accelerometer may not always be worn in exactly the same
way, a method for the adaptation to the user is performed. At
the beginning of each recording for each person there is
initialization (normalization) period of 15 seconds. The
average orientation angle of posture standing was measured as
φ
0
. The difference between the “ideal” standing orientation
angle (e.g. 180
o
for the x axis) and φ
0
was calculated as φ
diff
=
φ
ideal
φ
0
. After the normalization period, to each newly
calculated orientation angle φ
i
the difference angle φ
diff
is
added and finally the normalized angle is calculated as φ
normal
= φ
i
+ φ
diff
. This adaptation procedure is performed for each
axis for each accelerometer. Without this adaptation technique
the model and the results were person dependable. There was
a big difference in the accuracy for people wearing the
accelerometers in slightly different way.

A sliding window is used for calculation of the statistical
attributes. The current data sample and 5 past data samples are
combined in one window. The window size is chosen to be 6
data samples (1 second interval) because in this task there are
short-lasting activities that the system should detect (going
down, standing up). The first attribute from this group is the
Mean value of the data in the window. This attribute is
actually performing a low-pass filter to the raw data. The
filtered data is smoother and has fewer changes. This is a good
feature for the posture recognition process. The mathematical
definition for the x axis is:
n
a
Mean
i
x
n
i
x
1
. (3)
The number of data samples n is 6 (1 second window size).
The variable
i
x
a
is the acceleration along x axis. Using the
same formula, mean values for other axes Mean
y
and Mean
z
are calculated. Also the mean value for the length of the
acceleration vector Mean
length
is calculated. A similar approach
is used for the Standard deviation and the Root mean square
attributes.
The Root Mean Square is a similar attribute to the Mean
value, but it is useful when the observed value is varying
above and below zero. That is the case in our acceleration
values. Depending of the orientation of the accelerometer, the
values can be positive or negative (e.g. +g or g). The RMS
for the length of the acceleration vector is computed as
follows:
n
length
RMS
i
n
i
2
1
length
. (4)
The variable
2
i
le ngth
is the square of the length of the
acceleration vector for the current member in the sum.
Similarly the RMS
x
, RMS
y
and RMS
z
were calculated.
The Standard Deviation attribute is good for distinguishing
long-lasting static postures/activities from transitional
postures/activities. It can detect when the movement of the
accelerometer is intense. The mathematical definition for the
length of the acceleration vector is:
n
lengthlength
STD
i
n
i
2
1
length
)(
. (5)
The variable
i
length
is the length for the current member in
the sum and
is the mean value in the current window.
Also the standard deviation for each of the axes was
calculated: STD
x
, STD
y
and STD
z
.
When a person’s body is static, the accelerometers respond
only to the gravity, producing a constant 1 g total acceleration.
During motion the accelerometers produce changing
acceleration signal and the fiercer the motion, the greater the
change in the signal. Using these changes in the acceleration
vector an attribute is computed for the detection of sensor
movement - Acceleration Vector Changes (AVC). AVC value
of this attribute increases as the accelerometer is in movement
(walking, going down, standing up etc.). This attribute takes in
consideration the last 2 seconds of data (12 data samples). It
sums up the last 12 differences of lengths of the acceleration
vector and divides the sum with the time interval (2 seconds)
of the data. AVC is computed as follows:
0
1
1
0
||
TT
lengthlength
AVC
n
ii
n
i
. (6)
T
0
is the time stamp for the first data sample in the window
and T
n
is the time stamp of the last data sample. With this
attribute the movement of the person can be detected: it
distinguishes static from dynamic postures. A boolean
(true/false) attribute, which compares the AVC attribute value
to a threshold, is also computed. If the value is above the
threshold, the boolean attribute is true, otherwise it is false.
The threshold value is 0.0015 and it was chosen empirically
after series of tests on recordings different from the test ones.
All these 21 attributes were extracted for each
accelerometer and collected together in one attribute vector.
This vector was passed through to the classification model,
which tried to predict the appropriate posture. The
classification model was previously trained. Because we
recorded 11 people, the model was trained on 10 people and
tested on the remaining person. This procedure was repeated
for each person. The decision which classification algorithm to
be used was made after evaluating the results in the Weka
toolkit. Several commonly used classification algorithms were
analyzed: Naïve Bayes, SVM, J48 and Random Forest. The
algorithm that achieved the best results for almost all postures
was Random Forest.
B. Fall Detection Module
The second module is focused on the detection of fall
events. In this module two acceleration-based algorithms are
used. The first one detects high accelerations using only one
accelerometer (chest or waist). It is a threshold-based
algorithm and has some improvements for reducing false
alarms. The acceleration pattern during a typical fall is a
decrease in acceleration followed by a fast increase. This is
shown in Figure 2. The reason for this pattern is that the
acceleration at rest is 1 g and during free fall 0 g. When a
person starts falling, the acceleration decreases from 1 g to
around 0.5 g (perfect free fall is never achieved). Upon the
impact with the ground, a short increase in the acceleration is
measured.
Figure 2: Acceleration pattern during a fall.

To detect falls with a threshold, we used the length of the
acceleration vector, which means that we ignored the
orientation of the accelerometer. The first idea was to use a
simple threshold that will detect only the high acceleration
(impact). This resulted in false positives during quickly
standing up. The reason for this is that the quickly standing up
has also a high acceleration and can be confused with a fall.
However, the pattern during standing up is a reverse compared
to the fall pattern: first the increase is detected followed by the
decrease. Using this information the minimum and the
maximum acceleration within 1.5 second window were
measured. If the difference between the maximum and the
minimum exceeded the threshold and the maximum (impact)
appeared after the minimum (fall), we declared that a fall had
occurred. The threshold was chosen empirically for each of
the accelerometers individually. Eventually the waist
accelerometer had the threshold of 0.8 g and chest
accelerometer had 1 g. The reason for this is the placement of
the accelerometers on the body and the impact with the
ground. This method works perfectly on normal fast falls, and
the false alarms rate during normal activities is reduced to
minimum.
The problem with this algorithm appears if there is high
acceleration (decrease followed by increase) but the event is
not fall (e.g. quickly sitting on the chair). To solve this issue
we used a second algorithm, which takes into account the
recognized posture after a potential fall event. The improved
algorithm developed for detecting fall events used the postures
recognized from the posture recognition module as input. It
analyzed the recognized postures and decided if a fall was
detected. Two rules were implemented. The fall had occurred
if:
Acceleration has exceeded the threshold as described in the
first algorithm AND the person is lying for more than 10
seconds; OR
The person sits on the ground more than 10 seconds.
The assumption of the second rule, which is already implicit in
the scenario, was that the elderly usually do not sit on the
ground.
The decision about the posture of the person in 10 seconds
interval was done by choosing the major predicted posture in
this interval. In other words, the posture with highest number
of data samples was chosen.
V. EXPERIMENTAL RESULTS
We compared the performances of the posture recognition
and fall detection methods on a test scenario. It was recorded
by 11 healthy volunteers (7 male and 4 female), 5 times by
each person. We made a comparison to a location system. This
was possible because each person was wearing four
accelerometers and four location tags during the recordings.
A. Test Scenario
The test scenario is around 15 minutes long and includes
all the target body postures. It was designed specifically to
investigate events that may be difficult to recognize as falls or
non-falls. The events are listed in Table 1. They were recorded
in single recordings interspersed with short periods of
walking.
TABLE 1: EVENTS SEQUENCE IN THE TEST SCENARIO.
No.
Description
1
Sitting down normally on the chair
2
Tripping, falling fast on the ground
3
Lying down normally on the bed
4
Falling slowly (trying to hold onto furniture), lying
on the ground
5
Sitting down quickly on the chair
6
Falling from chair slowly when trying to stand up
(trying to hold onto furniture), landing sitting of
the ground
7
Sitting down quickly on the chair
8
Falling from chair quickly when trying to stand up,
landing sitting of the ground
9
Searching for something on the ground - on all
fours and lying
As shown in the section on related work, accelerometers
can accurately detect typical falls, so we included only one
such fall (event number 2) to demonstrate that the system can
recognize it accurately. We included two atypical falls (4, 6
and 8) to test the use of posture information, namely that a
person is not expected to sit on the ground (as opposed to the
chair). Furthermore, we included two events (5 and 7) that
involve high acceleration and could thus be misclassified as
falls by accelerometers. We also included an event (9) that
involves voluntary lying on the ground, which could mislead
the methods that use information other than acceleration. The
last two events (1 and 3) are perfectly normal and were
included to verify that all the methods work correctly and do
not recognize them as falls.
B. Results
The results are shown for each of the modules in separate
and different ways. For body posture recognition the results
are presented as the accuracy of the classifier for each of the
class values (postures). The confusion matrix for the waist
accelerometer is also shown. The results for fall detection are
shown in terms of events from the test scenario.
1) Body Posture Recognition
For body posture recognition module, leave one person out
technique was used for evaluating the results. That means the
model was trained on 10 people and tested on the remaining
person. This procedure was done 11 times, once for each
person. At the end the average accuracy was calculated for
each posture separately and the overall accuracy was
calculated for all the postures together. This evaluation
approach is more reliable than the ones that use separate
testing scenarios for each activity (only standing, only lying)
or that use the same person for training and testing. Using the
same person would give overly optimistic results if the
intended use of the model is to classify the postures of
previously unseen people. As we mentioned before every data
sample is classified. That means for one event of lying on the

Citations
More filters
Journal ArticleDOI
11 Dec 2015-Sensors
TL;DR: A review of different classification techniques used to recognize human activities from wearable inertial sensor data shows that the k-NN classifier provides the best performance compared to other supervised classification algorithms, whereas the HMM classifier is the one that gives the best results among unsupervised classification algorithms.
Abstract: This paper presents a review of different classification techniques used to recognize human activities from wearable inertial sensor data. Three inertial sensor units were used in this study and were worn by healthy subjects at key points of upper/lower body limbs (chest, right thigh and left ankle). Three main steps describe the activity recognition process: sensors' placement, data pre-processing and data classification. Four supervised classification techniques namely, k-Nearest Neighbor (k-NN), Support Vector Machines (SVM), Gaussian Mixture Models (GMM), and Random Forest (RF) as well as three unsupervised classification techniques namely, k-Means, Gaussian mixture models (GMM) and Hidden Markov Model (HMM), are compared in terms of correct classification rate, F-measure, recall, precision, and specificity. Raw data and extracted features are used separately as inputs of each classifier. The feature selection is performed using a wrapper approach based on the RF algorithm. Based on our experiments, the results obtained show that the k-NN classifier provides the best performance compared to other supervised classification algorithms, whereas the HMM classifier is the one that gives the best results among unsupervised classification algorithms. This comparison highlights which approach gives better performance in both supervised and unsupervised contexts. It should be noted that the obtained results are limited to the context of this study, which concerns the classification of the main daily living human activities using three wearable accelerometers placed at the chest, right shank and left ankle of the subject.

549 citations


Cites background from "Accelerometer Placement for Posture..."

  • ...Gjoreski, 2011 [25] Thigh, Waist, Chest, Ankle Lying, Sitting, Standing, All Fours, Transitional 91%...

    [...]

  • ...[25] studied the optimal location of accelerometers for fall detection....

    [...]

  • ...[25] studied the opti al locati f l ll i ....

    [...]

Journal ArticleDOI
17 Jul 2013-Sensors
TL;DR: Although data from all locations provided similar levels of accuracy, the hip was the best single location to record data for activity detection using a Support Vector Machine, providing small but significantly better accuracy than the other investigated locations.
Abstract: This article describes an investigation to determine the optimal placement of accelerometers for the purpose of detecting a range of everyday activities. The paper investigates the effect of combining data from accelerometers placed at various bodily locations on the accuracy of activity detection. Eight healthy males participated within the study. Data were collected from six wireless tri-axial accelerometers placed at the chest, wrist, lower back, hip, thigh and foot. Activities included walking, running on a motorized treadmill, sitting, lying, standing and walking up and down stairs. The Support Vector Machine provided the most accurate detection of activities of all the machine learning algorithms investigated. Although data from all locations provided similar levels of accuracy, the hip was the best single location to record data for activity detection using a Support Vector Machine, providing small but significantly better accuracy than the other investigated locations. Increasing the number of sensing locations from one to two or more statistically increased the accuracy of classification. There was no significant difference in accuracy when using two or more sensors. It was noted, however, that the difference in activity detection using single or multiple accelerometers may be more pronounced when trying to detect finer grain activities. Future work shall therefore investigate the effects of accelerometer placement on a larger range of these activities.

339 citations


Cites background from "Accelerometer Placement for Posture..."

  • ...[9] studied the best location to place accelerometers for fall detection, based on the classification of postures....

    [...]

  • ...Furthermore, the authors [9] did not report on the classification accuracy for all possible combinations of sensors....

    [...]

  • ...Gjoreski [9] Lying, sitting, standing, all fours, transitional (7) 11 Chest, Waist, Ankle, Thigh Orientation, Mean, Root Mean Square, Standard Deviation and Movement detection Random Forest (75%–99%)...

    [...]

Journal ArticleDOI
TL;DR: A smartphone-based fall detection system that monitors the movements of patients, recognizes a fall, and automatically sends a request for help to the caregivers, to reduce the problem of false alarms.

320 citations

Journal ArticleDOI
TL;DR: This work explores the problem of stress detection using machine learning and signal processing techniques in laboratory conditions, and then applies the extracted laboratory knowledge to real-life data to propose a novel context-based stress-detection method.

221 citations


Cites methods from "Accelerometer Placement for Posture..."

  • ...For this purpose, the Empatica device provides acceleration data, which has proven to be successful for recognizing activities [28,46]....

    [...]

Journal ArticleDOI
18 Jul 2014-Sensors
TL;DR: The aim of this study is to review the existing fall detection systems and some of the key research challenges faced by the research community in this field, and to categorize the existing platforms into two groups: wearable and ambient devices.
Abstract: Falls and fall-related injuries are major incidents, especially for elderly people, which often mark the onset of major deterioration of health. More than one-third of home-dwelling people aged 65 or above and two-thirds of those in residential care fall once or more each year. Reliable fall detection, as well as prevention, is an important research topic for monitoring elderly living alone in residential or hospital units. The aim of this study is to review the existing fall detection systems and some of the key research challenges faced by the research community in this field. We categorize the existing platforms into two groups: wearable and ambient devices; the classification methods are divided into rule-based and machine learning techniques. The relative merit and potential drawbacks are discussed, and we also outline some of the outstanding research challenges that emerging new platforms need to address.

209 citations


Cites background or methods from "Accelerometer Placement for Posture..."

  • ...[55] ● [75] ● ● ● ● ● [84] ● ● ● ● ● [52] ● ● ● ● [106] ● [69] ● ● ● ● ● ● [105] ● ● [79] ● ● ● ● [49] ● [70] ● ● ● ● ● ● ● ● [93] ● ● ● ● [53] ● ● ● ● [50] ● ● ● ● ● [98] ● ● ● [101] ● ● ● ● [76] ● ● ● ● ● ● ● ● ● ● ● ● ● [66] ● [46] ● ● ● ● ● ● ● ● [67] ● ● ● ● ● ● ● ● [48] ● ● ● [83] ● ● ● ● [85] ● ● ● ● [92] ● [80] ● [102] ● ● ● [77] ● ● ● ● [68] ● ● ● ● ● ● [104] ● [71] ● ● ● ● ● ● [86] ● ● [89] ● ● ● ● ● ● ● ● [88] ● ● ● ● ● ● ● ● [78] ● ● ● ● [41] ● [81] ● ● ● ● ● ● ● ● [103] ● ● ● ● [107] ● ● ● ● ● ● ● ● [99] ● [90] ● ● ● ● [82] ● ● ● [38] ● ● ● ● [35] ● ● ● [73] ● ● [96] ● ● ● ● [32] ● ● ●...

    [...]

  • ...(2011) [49] 4 4 F1, F2, F5, F9, F17, F29 TB, RB, DT, NB Acc = 94%...

    [...]

  • ...[49] included slow falls with an attempt to hold onto the furniture in their data collection protocol....

    [...]

  • ...To calculate the angle between the device and the gravitational vector, some studies used the Earth’s standard acceleration due to gravity (g) [67,104] or the sum vector magnitude (F5) equal to 1g (assuming no movement in a static activity) [46,49] in their equations....

    [...]

  • ...[49] compared sensors placed at the chest, waist, right ankle, and right thigh, and on the other hand, reported the waist as the optimal position....

    [...]

References
More filters
Proceedings Article
09 Jul 2005
TL;DR: This paper reports on the efforts to recognize user activity from accelerometer data and performance of base-level and meta-level classifiers, and Plurality Voting is found to perform consistently well across different settings.
Abstract: Activity recognition fits within the bigger framework of context awareness. In this paper, we report on our efforts to recognize user activity from accelerometer data. Activity recognition is formulated as a classification problem. Performance of base-level classifiers and meta-level classifiers is compared. Plurality Voting is found to perform consistently well across different settings.

1,561 citations


"Accelerometer Placement for Posture..." refers background in this paper

  • ...Using these coordinates for each of the tags, the system is able to reconstruct the user's posture and detect alarming situations [2]....

    [...]

Proceedings ArticleDOI
03 Jun 2009
TL;DR: A novel fall detection system using both accelerometers and gyroscopes that reduces both false positives and false negatives, while improving fall detection accuracy, and features low computational cost and real-time response.
Abstract: Falls are dangerous for the aged population as they can adversely affect health. Therefore, many fall detection systems have been developed. However, prevalent methods only use accelerometers to isolate falls from activities of daily living (ADL). This makes it difficult to distinguish real falls from certain fall-like activities such as sitting down quickly and jumping, resulting in many false positives. Body orientation is also used as a means of detecting falls, but it is not very useful when the ending position is not horizontal, e.g. falls happen on stairs. In this paper we present a novel fall detection system using both accelerometers and gyroscopes. We divide human activities into two categories: static postures and dynamic transitions. By using two tri-axial accelerometers at separate body locations, our system can recognize four kinds of static postures: standing, bending, sitting, and lying. Motions between these static postures are considered as dynamic transitions. Linear acceleration and angular velocity are measured to determine whether motion transitions are intentional. If the transition before a lying posture is not intentional, a fall event is detected. Our algorithm, coupled with accelerometers and gyroscopes, reduces both false positives and false negatives, while improving fall detection accuracy. In addition, our solution features low computational cost and real-time response.

543 citations


"Accelerometer Placement for Posture..." refers background in this paper

  • ...Using these coordinates for each of the tags, the system is able to reconstruct the user's posture and detect alarming situations [2]....

    [...]

Proceedings ArticleDOI
04 Dec 2009
TL;DR: A high-accuracy human activity recognition system based on single tri-axis accelerometer for use in a naturalistic environment that exploits the discrete cosine transform, the Principal Component Analysis (PCA) and Support Vector Machine for classification human different activity.
Abstract: This paper developed a high-accuracy human activity recognition system based on single tri-axis accelerometer for use in a naturalistic environment. This system exploits the discrete cosine transform (DCT), the Principal Component Analysis (PCA) and Support Vector Machine (SVM) for classification human different activity. First, the effective features are extracted from accelerometer data using DCT. Next, feature dimension is reduced by PCA in DCT domain. After implementing the PCA, the most invariant and discriminating information for recognition is maintained. As a consequence, Multi-class Support Vector Machines is adopted to distinguish different human activities. Experiment results show that the proposed system achieves the best accuracy is 97.51%, which is better than other approaches.

262 citations


"Accelerometer Placement for Posture..." refers background in this paper

  • ...Using these coordinates for each of the tags, the system is able to reconstruct the user's posture and detect alarming situations [2]....

    [...]

Book ChapterDOI
01 Jan 2006
TL;DR: The preliminary results show that this method can detect the falls effectively, and reduce the probability of being damaged in the experiments for the elderly people.
Abstract: The fall is a crucial problem in the elderly people’s daily life, and the early detection of fall is very important to rescue the subjects and avoid the badly prognosis. In this paper, we use a wearable tri-axial accelerometer to capture the movement data of human body, and propose a novel fall detection method based on one-class support vector machine (SVM). The one-class SVM model is trained by the positive samples from the falls of younger volunteers and a dummy, and the outliers from the non-fall daily activities of younger and the elderly volunteers. The preliminary results show that this method can detect the falls effectively, and reduce the probability of being damaged in the experiments for the elderly people.

225 citations


"Accelerometer Placement for Posture..." refers background in this paper

  • ...Using these coordinates for each of the tags, the system is able to reconstruct the user's posture and detect alarming situations [2]....

    [...]

Proceedings ArticleDOI
22 Oct 2007
TL;DR: Measurements from the waist and head have potential to distinguish between falls and ADL and when the simple threshold-based detection was combined with posture detection after the fall, the specificity and specificity of fall detection were up to 100 %.
Abstract: The increasing population of elderly people is mainly living in a home-dwelling environment and needs applications to support their independency and safety. Falls are one of the major health risks that affect the quality of life among older adults. Body attached accelerometers have been used to detect falls. The placement of the accelerometric sensor as well as the fall detection algorithms are still under investigation. The aim of the present pilot study was to determine acceleration thresholds for fall detection, using triaxial accelerometric measurements at the waist, wrist, and head. Intentional falls (forward, backward, and lateral) and activities of daily living (ADL) were performed by two voluntary subjects. The results showed that measurements from the waist and head have potential to distinguish between falls and ADL. Especially, when the simple threshold-based detection was combined with posture detection after the fall, the sensitivity and specificity of fall detection were up to 100 %. On the contrary, the wrist did not appear to be an optimal site for fall detection.

211 citations


"Accelerometer Placement for Posture..." refers background in this paper

  • ...Using these coordinates for each of the tags, the system is able to reconstruct the user's posture and detect alarming situations [2]....

    [...]

Frequently Asked Questions (1)
Q1. What are the contributions in "Accelerometer placement for posture recognition and fall detection" ?

This paper presents an approach to fall detection with accelerometers that exploits posture recognition to identify postures that may be the result of a fall. Posture recognition as a standalone task was also studied.