Informatica 33 (2009) 205–212 205

Fall Detection and Activity Recognition with Machine Learning
Mitja Luštrek and Boštjan Kaluža
Jožef Stefan Institute, Department of Intelligent Systems
Jamova cesta 39, SI-1000 Ljubljana, Slovenia
E-mail: mitja.lustrek@ijs.si, bostjan.kaluza@ijs.si

Keywords: fall detection, activity recognition, posture and movement reconstruction, machine learning
Received: July 16, 2008

Due to the rapid aging of the European population, an effort needs to be made to ensure that the elderly
can live longer independently with minimal support of the working-age population. The Confidence
project aims to do this by unobtrusively monitoring their activity to recognize falls and other health
problems. This is achieved by equipping the user with radio tags, from which the locations of body parts
are determined, thus enabling posture and movement reconstruction. In the paper we first give a general
overview of the research on fall detection and activity recognition. We proceed to describe the machine
learning approach to activity recognition to be used in the Confidence project. In this approach, the
attributes characterizing the user’s behavior and a machine learning algorithm must be selected. The
attributes we consider are the locations of body parts in the reference coordinate system (fixed with
respect to the environment), the locations of body parts in a body coordinate system (affixed to the
user’s body) and the angles between adjacent body parts. Eight machine learning algorithms are
compared. The highest classification accuracy of over 95 % is achieved by Support Vector Machine
used on the reference attributes and angles.
Povzetek: Članek opisuje zaznavanje padcev in prepoznavanja aktivnosti nasploh ter izvedbo
prepoznavanja aktivnosti s strojnim učenjem za potrebe projekta Confidence.

1 Introduction
The European population is aging due to the increase in
life expectancy and decrease in birth rate. The percentage
of population aged over 65 years is anticipated to rise
from 17.9 % in 2007 to 53.5 % in 2060 [7]. As a
consequence, the number of the elderly will exceed the
society’s capacity for taking care of them. Thus an effort
needs to be made to ensure that the elderly can live
longer independently with minimal support of the
working-age population. This is the primary goal of the
EC Seventh Framework project Confidence [4].
The Confidence project will develop a ubiquitous
care system to unobtrusively monitor the user, raise an
alarm if a fall is detected and warn of changes in
behavior that may indicate a health problem. This will
improve the chances of a timely medical intervention and
give the user a sense of security and confidence, thus
prolonging his/her independence.
The user of the Confidence system will wear small
inexpensive wireless tags on the significant places on the
body, such as wrists, elbows, shoulders, ankles, knees
and hips. The precise number and placement of tags will
be defined during development. The tags may even be
sewn into the clothes. The locations of the tags will be
detected by a base station placed in the apartment and a
portable device carried outside. This will make it
possible to reconstruct the user’s posture and movement
and to recognize his/her activity. Some tags may be
placed in the user’s environment at locations such as bed
and chair to recognize activities such as the user lying in
a bed and sitting in a chair. Finally, the user’s behavior
will be interpreted as normal or abnormal. An alarm or
warning will be raised in the latter case.
This paper describes machine learning methods for
activity recognition [12][13] to be used in the Confidence
project. We focus on the selection of attributes and
machine learning algorithm to maximize the recognition
accuracy. The activities to recognize are falling, the
process of lying down, the process of sitting down,
standing/walking, sitting and lying. Falling is important
in itself because fall detection is one of the main goals of
the project. For the processes of lying down and sitting
down, we wanted to see whether they can be
distinguished from falling. The recognition of
standing/walking, sitting and lying is needed to detect
changes in behavior, such as the user walking less and
lying more, which may indicate a health problem.
The paper is structured as follows. Section 2 gives a
detailed overview of related work on fall detection and
activity recognition [9]. Section 3 describes the
recordings of user behavior used as input data. Section 4
lists the attributes extracted from the input data that are
fed into the machine learning algorithms. Section 5
presents the experiments in which the various attributes
and machine learning algorithms are compared. Finally,
Section 6 concludes the paper in outlines the future work.

206 Informatica 33 (2009) 205–212 M. Luštrek et al.
2 Related work
We divide the work on fall detection and activity
recognition into four approaches presented in the
following four subsections. They are distinguished by the
equipment used and by the features extracted from sensor
data.
The first approach is based on accelerometers. An
accelerometer is a device for detecting the magnitude and
direction of the acceleration along a single axis or along
multiple axes. Three-axis accelerometers are typically
used. By detecting the acceleration caused by the earth’s
gravity, one can also compute the accelerometer’s angle
with respect to the earth.
The second approach uses gyroscopes, which
measure orientation. A gyroscope consists of a spinning
wheel whose axle is free to take any orientation. It can
measure the orientation along one axis or multiple axes.
By equipping an object with the gyroscope(s) to measure
the orientation along three axes, it is possible to exactly
determine the object’s orientation and the changes in
orientation, from which the angular velocity can be
computed.
The third approach is denoted visual detection
without posture reconstruction. It is based on extracting
input data from still images or from video. Various
computer vision techniques are applied to the input data,
but the human posture is not reconstructed explicitly.
The fourth approach, named visual detection with
posture reconstruction, is based on 3D locations of
markers placed on an object, typically human body. The
approach also uses video recordings, but, in contrast to
the third approach, the visual information is used only to
reconstruct the 3D locations of the markers. Additional
processing uses the markers’ coordinates as input data. If
a sufficient number of markers are provided, it is
possible to reconstruct the shape of an object, which in
our case means the human posture.
2.1 Accelerometers
The most common and simple methodology for fall
detection is using a tri-axial accelerometer with threshold
algorithms [3][10]. Such algorithms simply raise the
alarm when the threshold value of acceleration is
reached. There are several sensors with hardware built-in
fall detection [1][5][15], having the accuracy of over
80 %.
Zhang et al. [25] designed a fall detector based on
Support Vector Machine (SVM) algorithm. The detector
was using one waist-worn accelerometer. The features
for machine learning were the accelerations in each
direction, changes in acceleration etc. Their method
detected falls with 96.7 % accuracy. Researches
embedded an accelerometer in a cell phone [24] and
detected falls with the proposed method. The cell phone
was put in a pocket of clothes or hanged around the neck,
which made the detection more difficult as with the
body-fixed sensor. The cell-phone system correctly
raised the alarm in 93.3 % of the cases.
Tapia et al. [18] presented a real-time algorithm for
automatic recognition of not only physical activities, but
also, in some cases, their intensities, using five wireless
accelerometers and a wireless heart rate monitor. The
accelerometers were placed at shoulder, wrist, hip, upper
part of the thigh and ankle. The features, e.g., FFT peaks,
variance, energy, correlation coefficients, were extracted
from time and frequency domains using a predefined
window size on the signal. The classification of activity
was done with C4.5 and Naïve Bayes classifiers into
three groups: postures (standing, sitting etc.), activities
(walking, cycling etc.) and other activities (running,
using stairs etc). For these three classes they obtained the
recognition accuracy of 94.6 % using subject-dependent
training and 56.3 % using subject-independent training.
Willis [21] developed a fall detection system based
on belief network models, which enable probabilistic
modeling of scenarios (e.g., normal walking,
tripping/stumbling and running) and the transitions
between them. The sensors were placed under the heel
and toe, which made it possible to reconstruct gait cycle
and to detect falls. The accuracy was not reported.
Researchers using accelerometers give a lot of
attention to the optimal sensor placement on the body
[3][10]. A head-worn accelerometer provides excellent
impact detection sensitivity, but its limitations are
usability and user acceptance. A better option is a waist-
worn accelerometer. The wrist did not appear to be an
optimal site for fall detection. Some researchers made a
step further and used accelerometers for trying to
recognize the impact and posture after the fall [11].
In the Confidence system, accelerations could in
principle be derived from the movement of tags.
However, we believe this approach to be unreliable: first,
because the acceleration is the second derivative of tag
location and as such strongly affected by sensor noise,
and second, because the data acquisition frequency in
Confidence is expected to be relatively low. The studies
of sensor placement may be valuable for deciding where
to place tags in Confidence.
2.2 Gyroscopes
Bourke and Lyons [2] introduced a threshold algorithm
to distinguish between normal activities (sitting down
and standing up, lying down and standing up, getting in
and out of a car seat, walking etc.) and falls. The ability
to discriminate was achieved using a bi-axial gyroscope
mounted on the torso, measuring pitch and roll angular
velocities. They applied a threshold algorithm to the
peaks in the angular velocity signal, angular acceleration
and torso angle change. The system proved 100 %
successful in fall detection.
The Confidence system derives velocities from the
movement of tags. The velocity, being the second
derivative of tag location and being less affected by the
low data acquisition frequency, is more reliable than
acceleration. However, since the data available in
Confidence is much richer than that provided by
gyroscopes, we decided against simple threshold-based
fall detection.

FALL DETECTION AND ACTIVITY... Informatica 33 (2009) 205–212 207
2.3 Visual detection without posture
reconstruction
Vishwakarma et al. [20] presented a video approach for
fall detection. First, they eliminated the background of
the video and extracted a set of features from the
remaining objects’ bounding boxes, e.g., the aspect ratio,
horizontal and vertical gradients etc. In the next step they
detected falls based on the angle between an object’s
bounding box and the ground. The final step was fall
confirmation, which was rule-based, e.g., the
abovementioned angle had to be less than 45°. The
method achieved 95 % accuracy on single-object fall
detection and 64 % accuracy on multiple objects.
Fu et al. [8] described a vision system designed to
detect accidental falls in elderly home care applications.
They used a temporal contrast vision sensor, which
extracts changing pixels from the background. An
algorithm was observing the dynamic of motion and
reported falls when it indicated significant changes in the
vertical downward direction. They were able to
distinguish falls from normal human behaviors, such as
walking, crouching down and sitting down. The accuracy
was not reported.
The proposed methods are quite capable of dealing
with fall detection, but it is not clear how to adapt them
to the sensor data available in the Confidence system.
2.4 Visual detection with posture
reconstruction
Wu [23] studied unique features of the velocity during
normal and abnormal (i.e. fall) activities so as to make
the automatic detection of falls during the descending
phase of a fall possible. Normal activities included
walking, rising from a chair and sitting down, descending
stairs, picking up an object from the floor, transferring in
and out of a tub and lying down on a bed. The study
provides exhaustive velocity parameters for fall
detection, gathered by three markers placed on the
posterior side of the torso, recorded by three cameras
with the sampling rate of 50 Hz. The aim of the study
was to suggest velocity characteristics, so the author did
not actually implement automatic fall detection.
Qian et al. [16] introduced a gesture-driven
interactive dance system capable of real-time feedback.
They used 41 markers on the body recorded by 8
cameras with the frame rate of 120 Hz to construct a
human body model. The model was used to extract
features such as torso orientation, angles between
adjacent body parts etc., which was used to represent
different gestures. Each gesture was statistically modeled
with a Gaussian random vector defined as the statistical
distribution of the features for that gesture. To recognize
a new pose, the likelihood of its feature vector given the
vector of each known gesture was computed. The new
pose was classified as the gesture for which this
likelihood was the largest. Experimental results with two
dancers performing 21 different gestures achieved
gesture recognition rate of 99.3 %.
Sukthankar and Sycara [17] presented a system that
reconstructs the users’ posture and recognizes pre-
defined behaviors. The data were captured with 43 body
markers and 12 cameras with the sampling rate of
120 Hz. They constructed a human body model from the
raw marker coordinates, and computed features, e.g. the
angles between body parts, limb lengths, range of motion
etc. from the model. Learning was performed using
SVM. The method achieved 76.9 % accuracy in
detecting the following elementary activities: walking,
running, sneaking, being wounded, probing, crouching,
and rising. Behavior was defined as a sequence of
elementary activities and was modeled with Hidden
Markov models. The authors defined a number of
behavior models and classified a new sequence of
activities into the model that fit it best.
The markers in the proposed systems have the same
role as the tags in the Confidence system. The methods
by Qian et al. and even more so by Sukthankar and
Sycara inspired the approach we used for activity
recognition in Confidence. We are not aware of anybody
having used this kind of methods for fall detection,
though.
3 Input data
The goal of our research was to classify the user’s
behavior into one of the following activities: falling,
lying down, sitting down, standing/walking, sitting and
lying. To obtain training data for a classifier to recognize
these activities, we recorded 45 examples of the behavior
of three persons. Each recording consisted of multiple
activities:
3 × 15 recordings of falling, consisting of
standing/walking, falling and lying.
3 × 10 recordings of lying down, consisting of
standing/walking, lying down and lying.
3 × 10 recordings of sitting down, consisting of
walking, sitting down and sitting.
3 × 10 recordings of walking.
The recordings consisted of the coordinates of 12
body tags attached to the shoulders, elbows, wrists, hips,
knees and ankles. This is the full complement of tags that
will probably be reduced in the future. Since the
equipment with which the Confidence system will
acquire tag coordinates is still under development, the
commercially available Smart infrared motion capture
system [6] was used instead. The coordinates were
acquired with 60 Hz. The frequency was afterwards
reduced to 10 Hz, which is the expected Confidence data
acquisition frequency. To make the recordings even more
similar to what we expect of the Confidence equipment,
we added Gaussian noise to them. The standard deviation
of the noise was 4.36 cm horizontally and 5.44 cm
vertically. This corresponds to the noise measured in the
Ubisense real time location system [19]. The Ubisense
system is similar to the equipment planned for acquiring
tag coordinates in Confidence. The noise in the
recordings was smoothed with Kalman filter [14].

208 Informatica 33 (2009) 205–212 M. Luštrek et al.
4 Attributes for machine learning
Finding the appropriate representation of the user’s
behavior activity was probably the most challenging part
of our research. The behavior needs to be represented
with simple and general attributes, so that the classifier
using these attributes will also be general and work well
on behaviors different from those in our recordings. It is
not difficult to design attributes specific to our
recordings; such attributes would work well on them.
However, since our recordings captured only a small part
of the whole range of human behavior, overly specific
attributes would likely fail on general behavior.
The attribute vector from which the classifier infers
the user’s activity consists of ten consecutive snapshots
of the user’s posture, describing one second of activity.
When multiple activities took place within a given
second, the attribute vector was assigned the longest one.
We designed three sets of attributes describing the
user’s behavior. Reference attributes are expressed in the
reference coordinate system, which is fixed with respect
to the user’s environment. Body attributes are expressed
in a coordinate system affixed to the user’s body. Angle
attributes are the angles between adjacent body parts.
4.1 Reference attributes
When selecting reference attributes, we ignored x and y
coordinates. These coordinates describe the user’s
location in the environment, but the activities of interest
can generally take place at any location.
In the list of reference attributes, the upper index t
indicates the time within the one-second interval: t = 1 ...
10. The lower index i indicates the tag: i = 1 ... 12. The
lower index R indicates the reference coordinate system
and distinguishes reference attributes from those
belonging to the other two sets.
ztiR … z coordinate of tag i at time t
vtiR … the absolute velocity of the tag
vtziR … the velocity of the tag in the z direction
dtijR ... the absolute distance between the tags i
and j; j = i + 1 ... 12
dtzijR ... the distance between tags i in j in the z
direction
4.2 Body attributes
Body attributes are expressed in a coordinate system
affixed to the user’s body. This makes it possible to
observe x and y coordinates of the user’s body parts,
since these coordinates no longer depend on the user’s
location in the environment.
The body coordinate system is shown in Figure 1. Its
origin O is at the mid-point of the line connecting the hip
tags (HR and HL for the right and left hip respectively).
This line also defines the y axis, which points towards
the left hip. The z axis is perpendicular to the y axis,
touches the line connecting both shoulder tags (SR and SL
for the right and left shoulder respectively) at point Sz,
and points upwards. The x axis is perpendicular to the y
and z axes and points forwards.

Figure 1: The body coordinate system.
In order to translate reference coordinates into body
coordinates, we need to express the origin O and basis (i,
j, k) of the body coordinate system in the reference
coordinate system. Note that bold type denotes vectors
and x denotes a vector from the origin to the point X.
Equation (1) expresses the origin of the body coordinate
system in the reference coordinate system.
(1)
Equation (2) gives us the basis vector j.
(2)
To obtain k, Equation (3) is first used to calculate sz.



(3)
Once sz is calculated, Equation (4) gives us k.
(4)
Finally we obtain i using Equation (5).
(5)
We also experimented with a variant of body
coordinate system with the reference z axis, which is
shown in Figure 2. Its origin O is again at the mid-point
of the line connecting the hip tags. The z axis is the z
axis of the reference coordinate system. The y axis is
perpendicular to the z axis, lies on the plane defined by
the hip tags and a point on the z axis, and points towards
the left hip. The x axis is perpendicular to the y and z
axes and points forwards when the user is upright (in
general it points in the direction of the cross product of
the basis vectors j and k).

FALL DETECTION AND ACTIVITY... Informatica 33 (2009) 205–212 209

Figure 2: The body coordinate system with reference z
axis.
In the body coordinate system with the reference z
axis, the origin is again calculated with Equation (1). The
basis vector k equals the basis vector k in the reference
coordinate system: k = (0, 0, 1). The basis vector i is
perpendicular to k and to the vector from O to HL, which
is expressed with Equation (6).
(6)
The basis vector j is obtained with Equation (7).
(7)
To finally translate the coordinates in the reference
coordinate system into the coordinates in either of the
body coordinate systems, Equation (8) is used. The
vector pR = (xR, yR, zR, 1) corresponds to the point (xR, yR,
zR) in the reference coordinate system. The vector pB =
(xB, yB, zB, 1) corresponds to the point (xB, yB, zB) in a
body coordinate system. TR→B is the transformation
matrix from the reference to the body coordinate system.
Notation i(B)R refers to the basis vector i belonging to the
body coordinate system, expressed in the reference
coordinate system.


(8)
Body attributes (in either of the body coordinate
systems) are labeled with a lower index B:
(xtiB, y
t
iB, z
t
iB) ... coordinates of the tag i at the
time t
vtiB … absolute velocity of the tag
(φtiB, θ
t
iB) … the angles of movement of the tag
with respect to the z axis and xz plane
If a body coordinate system is used, the attributes
describing its location, orientation and movement with
respect to the reference coordinate system are added to
the attribute vector:
ztOR ... z coordinate of the origin of the body
coordinate system
(ΦtOR, Θ
t
OR) ... the direction of the x axis of the
body coordinate system with respect to the z
axis and xz plane
vtOR … absolute velocity of the origin of the
body coordinate system
(φtOR, θ
t
OR) ... the angles of movement of the
origin of the body coordinate system with
respect to the z axis and xz plane
So far we expressed body attributes in the body
coordinate system of each snapshot of the user's posture.
However, the attributes in all ten snapshots within a one-
second interval can be expressed in the coordinate
system belonging to the first snapshot in the interval.
This captures the changes in the x and y coordinates
between snapshots within the interval. First-snapshot
body attributes are the same as body attributes, except
that they are labeled with Bf instead of B. The attributes
describing the location and orientation of the first-
snapshot body coordinate system with respect to the
reference coordinate system are somewhat different,
though:
zOfR … z coordinate of the origin of the first-
snapshot body coordinate system
(ΦOfR, ΘOfR) … the direction of the x axis of the
first-snapshot body coordinate system with
respect to the z axis and xz plane
4.3 Angle attributes
The paper will not delve into the details of the
computation of body angles. The angles between body
parts that rotate in more than one direction are expressed
with quaternions:
qtSL and q
t
SR ... left and right shoulder angles
with respect to the upper torso at the time t
qtHL and q
t
HR ... left and right hip angles with
respect to the lower torso
qtT ... the angle between the lower and upper
torso
αtEL, α
t
ER, α
t
KL and α
t
KR ... left and right elbow
angles, left and right knee angles
5 Machine learning experiments
We tried various machine learning algorithms to train
classifiers for classifying the behavior into the six
activities (falling, lying down, sitting down,
standing/walking, sitting and lying). To do so, sections of
the 135 recordings described in Section 3 were first
manually labeled with the activities. Afterwards, the
recordings were split into overlapping one-second
intervals (one interval starting every one-tenth of a
second). The attributes described in Section 4 were
extracted from these intervals. This gave us 5,760
attribute vectors consisting of 240–2,700 attributes each
(depending on the combination of attributes used). An
activity was then assigned to each attribute vector.
Finally these vectors were used as training data for eight
machine learning algorithms: C4.5 decision trees,
RIPPER decision rules, Naive Bayes, 3-Nearest
Neighbors, Support Vector Machine (SVM), Random

210 Informatica 33 (2009) 205–212 M. Luštrek et al.
Forest, Bagging and Adaboost M1 boosting. The
algorithms were implemented in Weka [22], an open-
source machine learning suite. Default parameter settings
were used in all cases, except for Adaboost M1, where
the algorithm to train the base classifier was replaced
with Fast Decision Tree Learner. Machine learning
experiments proceeded in two steps.
In the first step of machine learning experiments we
compared the classification accuracy of the eight
machine learning algorithms and of all single attributes
sets described Section 4: reference, body, body with
reference z, first-snapshot body, first-snapshot body with
reference z and angles. The results are shown in Table 1.
The accuracy was computed with ten-fold cross-
validation. The accuracy of the best attribute set for each
algorithm is in bold type; the accuracy of the best
algorithm for each attribute set is on gray background.
Attribute set






Algorithm
re
fe
re
n
ce

b
o
d
y

b
o
d
y
w
it
h
r
ef
er
en
ce
z

fi
rs
t-
sn
ap
sh
o
t
b
o
d
y

fi
rs
t-
sn
ap
sh
o
t
b
o
d
y

w
it
h
r
ef
er
en
ce
z

an
g
le
s
Clean data
C4.5 decision trees 94.1 92.8 93.7 92.9 93.2 91.8
RIPPER
decision rules
93.1 91.4 92.8 92.0 93.0 90.9
Naive Bayes 89.5 88.7 90.6 86.8 88.2 76.7
3-Nearest Neighbor 97.1 92.0 82.8 88.1 85.1 96.9
SVM 97.7 94.4 95.0 94.1 94.3 90.5
Random Forest 97.0 96.5 96.8 96.0 96.0 96.8
Bagging 95.9 95.3 95.7 95.4 94.9 94.5
Adaboost M1
boosting
97.7 94.9 95.3 94.7 94.7 94.4
Noisy data
C4.5 decision trees 90.1 88.4 89.9 88.9 90.0 80.8
RIPPER
decision rules
87.5 84.7 88.1 86.2 88.6 80.0
Naive Bayes 83.9 79.1 84.0 81.0 82.2 78.2
3-Nearest Neighbor 95.3 74.6 79.7 73.4 74.7 93.3
SVM 96.3 87.2 91.6 89.9 91.1 87.2
Random Forest 93.9 90.5 93.4 91.9 93.2 90.5
Bagging 93.6 91.8 93.3 92.3 93.5 89.1
Adaboost M1
boosting
93.2 92.0 93.1 92.1 92.9 88.4
Table 1: Classification accuracy for all the algorithms
and all single attribute sets.
For the next step of machine learning experiments,
we retained the best algorithms and the best attribute
sets. To rank them, we compared the classification
accuracies of all pairs of algorithms and all pairs of
attribute sets. Table 2 shows the number of comparisons
in which a given algorithm statistically significantly (p <
0.05) wins over another algorithm, minus the number of
comparisons where it loses. Table 3 shows the same for
the attribute sets. The accuracies of the algorithms and
attribute sets selected for the second step are on grey
background; the accuracies of the best algorithm and
attribute set are in bold type. Since the second step
consisted of combining the attribute sets, the selection of
the sets to retain was based more on redundancy than
classification accuracy. Thus we retained angles, but not
the two first-snapshot body attributes (even though the
latter have a higher accuracy), because first-snapshot
body attributes are very similar to regular body attributes.
We chose body attributes with the body z axis over body
attributes with the reference z axis (even though the latter
again have a higher accuracy), because the reference z
coordinates are already included in the reference
attributes. The comparison between every-snapshot and
first-snapshot body attributes slightly favors the latter,
but we nevertheless retained the former because they are
computed more quickly.
Algorithm Wins – losses
Clean Noisy
C4.5 decision trees –12 –10
RIPPER decision rules –18 –21
Naive Bayes –38 –34
3-Nearest Neighbor –13 –16
SVM 13 11
Random Forest 38 23
Bagging 17 25
Adaboost M1 boosting 13 22
Table 2: The number of wins – the number of losses of
every algorithm against the others for clean and noisy
data
Attribute set Wins – losses
Clean Noisy
reference 25 28
body –2 –21
body with reference z 9 20
first-snapshot body –11 –9
first-snapshot body with reference z –2 12
angles –19 –30
Table 3: The number of wins – the number of losses of
every single attribute set against the others for clean and
noisy data
After selecting the best algorithms and attribute sets,
we proceeded with the second step of machine learning
experiments. In this step we tried combinations of
attribute sets. Table 4 shows the classification accuracy
for the four algorithms we retained and all the reasonable
combinations of the remaining attribute sets. The
accuracy of the best combination of attributes for each
algorithm is in bold type; the accuracy of the best
algorithm for each combination of attributes is on gray
background.

FALL DETECTION AND ACTIVITY... Informatica 33 (2009) 205–212 211
Attribute set
combination






Algorithm re
fe
re
n
ce
+
b
o
d
y

re
fe
re
n
ce
+

b
o
d
y
w
it
h
r
ef
er
en
ce
z

re
fe
re
n
ce
+
a
n
g
le
s
b
o
d
y
+
a
n
g
le
s
b
o
d
y
w
it
h
r
ef
er
en
ce
z
+

an
g
le
s
al
l
al
l
(r
ef
er
en
ce
z
)
Clean data
SVM 96.6 96.9 97.7 95.3 95.5 96.7 96.9
Random Forest 97.0 97.0 97.2 96.7 96.9 97.1 97.0
Bagging 96.1 96.0 96.1 95.6 95.7 96.3 96.0
Adaboost M1
boosting
95.7 95.6 95.5 95.3 95.3 95.6 95.5
Noisy data
SVM 95.5 95.4 96.5 91.9 92.5 95.6 95.5
Random Forest 93.8 94.2 94.1 91.8 93.5 93.9 94.0
Bagging 93.8 94.1 93.7 92.4 93.4 93.8 94.1
Adaboost M1
boosting
93.6 93.7 93.2 93.2 93.3 93.6 93.7
Table 4: Classification accuracy for the retained
algorithms and combinations of attribute sets.
6 Conclusion
We first investigated the work done so far in the area of
fall detection and activity recognition. Fall detection
methods were based on the accelerations and velocities
of body parts and on visual cues. These data will not be
available in the Confidence system, at least not directly.
What will be available are the locations of body parts.
Accelerations and velocities can be computed from the
changes in these locations, but with questionable
accuracy. We decided to use velocities, since they are
expected to be more accurate than accelerations, and the
locations of body parts themselves. Some work on
activity recognition was also based on accelerations and
velocities, but there were approaches better suited to
Confidence as well. We were mostly inspired by the
work of Sukthankar and Sycara [17], who used machine
learning on attributes representing the body posture.
We then examined various attributes and machine
learning algorithms to detect six common activities. The
attributes were the coordinates of body parts in the
reference coordinate system, the coordinates of body
parts in four different body coordinate systems and the
angles between adjacent body parts. We first compared
the attribute sets in isolation and then in combinations.
The reference coordinates were the best single attribute
set. In combination with the angles, they gave the highest
overall classification accuracy, although it should be
noted that all the combinations were close in
performance. We compared eight machine learning
algorithms, from which Support Vector Machine
produced the most accurate classifier: the accuracy on
clean data was 97.7 % and on noisy data 96.5 %. It was
closely followed by Random Forest, Bagging and
Adaboost M1 boosting.
There are four directions for future work. The first is
tuning the machine learning algorithms discussed in this
paper and augment them with feature selection
techniques. This is done relatively easily, but will
probably not contribute much to the classification
accuracy. The second direction is to take into account the
temporal information: each activity takes usually lasts for
some time and some transitions between activities are
more likely than others. This information can help us
correct some erroneous classifications, e.g., a single
falling in a long sequence of walking must be an error.
The third direction is using fewer than 12 tags, since a
potential product resulting from the Confidence project is
unlikely to use the full complement of tags. The last
direction for future work is experimenting with
recordings of additional behaviors. These may be
variations of the existing ones to test the robustness of
the classifier or entirely new activities to increase the
classifier’s scope.
Concerning the Confidence project, the results
described in this paper are encouraging. The
classification accuracy of over 95 % leads us to believe
that once the planned improvements are implemented,
the frequency of false alarms will be low enough for the
Confidence system to be useful.
Acknowledgement
This work was supported by the Slovenian Research
Agency under the Research Programme P2-0209
Artificial Intelligence and Intelligent Systems. The
research leading to these results has also received
funding from the European Community's Framework
Programme FP7/2007–2013 under grant agreement nº
214986. Consortium: CEIT (coordinator), Fraunhofer
Institute for Integrated Circuits (IIS), Jožef Stefan
Institute, Ikerlan, COOSS Marche, University of
Jyväskylä, Umeå Municipality, eDevice, CUP2000
S.p.A/Ltd., ZENON S.A. Robotics & Informatics. We
would like to thank Matjaž Gams for suggestions and
discussion and Barbara Tvrdi for help with
programming.
References
[1] AlertOne Services, Inc. iLife™ Fall Detection
Sensor. http://www.falldetection.com, 2008-07-18.
[2] Bourke, A. K., and Lyons, G. M. A threshold-based
fall-detection algorithm using a bi-axial gyroscope
sensor. Medical Engineering & Physics, vol. 30,
issue 1, pp. 84–90, 2006.
[3] Bourke, A. K., Scanaill, C. N., Culhane, K. M.,
O'Brien, J. V., and Lyons, G. M. An optimum
accelerometer configuration and simple algorithm
for accurately detecting falls. In Proceedings of the
24th IASTED international Conference on
Biomedical Engineering, pp. 156–160, 2006.
[4] Confidence. http://www.confidence-eu.org, 2008-
09-15

212 Informatica 33 (2009) 205–212 M. Luštrek et al.
[5] Doughty, K., Lewis, R., and McIntosh, A. The
design of a practical and reliable fall detector for
community and institutional telecare. Journal of
Telemedicine and Telecare, vol. 6, pp. 150–154,
2000.
[6] eMotion. Smart motion capture system. http://www.
emotion3d.com/smart/smart.html, 2008-09-15.
[7] Eurostat. http://epp.eurostat.ec.europa.eu, 2008-09-
09.
[8] Fu, Z., Culurciello, E., Lichtsteiner, P., and
Delbruck, T. Fall detection using an address-event
temporal contrast vision sensor. In Proceedengs of
the IEEE International Symposium on Circuits and
Systems – ISCAS 2008, pp. 424–427, 2008.
[9] Kaluža, B., and Luštrek M.. Fall Detection and
Activity Recognition Methods for the Confidence
Project: A Survey. In Proceedings of the 12th
International Multiconference Information Society
2008, vol. A, pp. 22–25, 2008.
[10] Kangas, M., Konttila, A., Lindgren, P., Winblad, P.,
and Jamsa, T. Comparison of low-complexity fall
detection algorithms for body attached
accelerometers. Gait & Posture, vol. 28, issue 2, pp.
285–291, 2008.
[11] Kangas, M., Konttila, A., Winblad, I., and Jamsa, T.
Determination of simple thresholds for
accelerometry-based parameters for fall detection.
In Proceedings of the 29th Annual International
Conference of the IEEE, Engineering in Medicine
and Biology Society, pp. 1367–1370, 2007.
[12] Luštrek, M., and Gams, M. Posture and movement
recognition from locations of body tags. European
Conference on Ambient Intelligence, "Ambient
Assisted Living" and "Personal Health" – between
Paragdigms, Projects and Products workshop, 2008.
[13] Luštrek, M., and Gams, M. Prepoznava položaja
telesa s strojnim učenjem. In Proceedings of the
12th International Multiconference Information
Society 2008, vol. A, pp. 30–33, 2008.
[14] Maybeck, P. S. Stochastic models, estimation, and
control. Mathematics in Science and Engineering
141, 1979.
[15] Noury, N., Barralon, P., Virone, G., Boissy, P.,
Hamel. M., and Rumeau, P. A smart sensor based
on rules and its evaluation in daily routines. In
Proceedings of the 25th Annual International
Conference of the IEEE, Engineering in Medicine
and Biology Society, vol. 4, pp. 3286–3289, 2003.
[16] Qian, G., Guo, F., Ingalls, T., Olson, L., James, J.,
and Rikakis, T. A Gesture-Driven Multimodal
Interactive Dance System. In Proceedings of the
International Conference on Multimedia and Expo,
Taipei, Taiwan, 2004.
[17] Sukthankar, G., and Sycara, K. A Cost Minimization
Approach to Human Behavior Recognition. In
Proceedings of the Fourth international Joint
Conference on Autonomous Agents and Multiagent
Systems 2005, pp. 1067–1074, 2005.
[18] Tapia, E. M., Intille, S. S., Haskell, W., Larson, K.,
Wright, J., King, A., and Friedman, R. Real-Time
Recognition of Physical Activities and Their
Intensities Using Wireless Accelerometers and a
Heart Rate Monitor. In Proceedengs of the 11th
IEEE International Symposium on Wearable
Computers, pp. 37–40, 2007.
[19] Ubisense. http://www.ubisense.net, 2008-09-15.
[20] Vishwakarma, V., Mandal, C., and Sura, S.
Automatic Detection of Human Fall in Video.
Pattern Recognition and Machine Intelligence:
Automatic Detection of Human Fall in Video, pp.
616–623, 2007.
[21] Willis, D. J. Ambulation Monitoring and Fall
Detection System using Dynamic Belief Networks.
PhD Thesis, School of Computer Science and
Software Engineering, Monash University, 2000.
[22] Witten, I. H., and Frank, E. Data Mining: Practical
machine learning tools and techniques, 2nd Edition.
Morgan Kaufmann, San Francisco, USA, 2005.
[23] Wu, G. Distinguishing fall activities from normal
activities by velocity characteristics. Journal of
Biomechanics, vol. 33, issue 11, pp. 1497–1500,
2000.
[24] Zhang, T., Wang, J., Liu, P., and Hou, J. Fall
Detection by Embedding an Accelerometer in
Cellphone and Using KFD Algorithm. International
Journal of Computer Science and Network Security,
vol. 6, issue 10, 2006.
[25] Zhang, T., Wang, J., Liu, P., and Hou, J.. Fall
Detection by Wearable Sensor and One-Class SVM
Algorithm. Lecture Notes in Control and
Information Science, issue 345, pp. 858–863, 2006.