![](bg2.png)
• Defining the technical characteristics of the
sensor needed to implement a system using the
selected algorithm.
• Implementing the algorithm on a prototype to
check its performance in simulated falls and
rapid movements (to eliminate false positives),
leading finally to real situations.
The rest of the article details the proposed solution,
beginning with the study of the detection algorithms
published in the medical literature. As a result, an
algorithm that combines the advantages of various
methods is proposed and the requirements of the
sensors needed for the implementation are stated.
Section 3 shows how a low cost accelerometer can
achieve those requirements and finally Section 4
presents the design and the results of the
simulations carried out. Finally, similar products are
compared, and the conclusions and future work
close the paper.
2 FALL DETECTION
An initial review of the literature convinced us of
the advantages of accelerometers as the most
suitable type of sensors to detect falls. Although
there are other alternatives, such as the use of
gyroscopes (Bourke and Lyons, 2008), most works
use two or three axes accelerometers (Bourke et al.,
2004) (Chen et al., 2005). To design a reliable
detection system based on these devices, the
accelerations naturally present in the human body
must be previously documented, both in normal
movements and different types of falls. Various
medical articles have studied these accelerations.
When a person falls and hits the ground, his body
suffers accelerations above those that occur when
he is performing a normal activity. The work (Chen
et al., 2005) studied the differences between sitting
movements and falls by means of experiments with
two two-axis accelerometers. Although the graphics
were very similar, during a typical fall the
acceleration is 7g, while the accelerations measured
when a person sits down are less than 3g (about
2.6g where measured). Looking at the graphs
presented in that article, it is noteworthy that, at the
beginning of the fall, acceleration decreases
(indicating the period of fall), but immediately there
is a large peak indicating the impact against the
ground (7g approx.). The accelerometer
measurements, before and after the fall, are held at
about 1g, as expected. Similar results, even with
major peaks, were observed in lateral falls.
From the viewpoint of the type of falls, Lord et
al. (Lord et al., 1993) found that 82% occurred
when people were upright. The most common falls
occurred while an elderly person is walking, slides
and falls. Another study, conducted by (O'Neill et
al., 1994), found that, of 180 crashes recorded, 160
were forward and, in 60% of these, the subject was
taking a step forward with one bent knee and one
foot in the air, the typical movement of a walking
step.
With these studies as a reference, (Bourke et al.,
2007) attempted to define the acceleration threshold
that can automatically discriminate between normal
body movements and different types of falls. The
values of the accelerations were derived from daily
activities performed by elderly people and
simulated falls performed by young people. The
first experiment involved ten elderly people, aged
between 70 and 83, with a tri-axial accelerometer,
placed first on the trunk and then on the thigh. The
activities were sitting and rising from an armchair
or a kitchen chair, walking 10 m, etc. The second
experiment used ten young people aged between 21
and 29 who simulated six different types of falls.
Although forward falls are more frequent, they also
simulated lateral falls, as these often produce a great
impact on the trunk and often result in fractures
when they happen. The authors selected the lowest
value of the accelerations recorded during simulated
falls (upper fall threshold), and the largest of the
smaller peaks (lower fall threshold). The smallest
accelerations during a fall were about 3.5g but
others were much greater, while normal activities
usually produced accelerations of 1 to 2.5g,
although sometimes there are activities, such as
running or sitting, which can surpass this. In
conclusion, the threshold of normal movements
should be between 0.41g and 3.52g. The
acceleration values outside this range could be
considered potential falls. The success percentage
of the algorithm, including false positives, was
calculated with the accelerometer placed on the
trunk and on the thigh. The best results were
obtained for the trunk, with more than 90% correct
hits. But false positives (false alarms) remained the
real problem.
(Chen et al., 2005) used a different approach
that took into account the unexpected changes in
body orientation. They also studied the situations of
repeated impacts to determine certain types of falls
(on staircases, for example) that may be especially
dangerous. Based on these studies, we propose an
experiment using an algorithm that combines the
orientation changes postulated by Chen and the
ICSOFT 2010 - 5th International Conference on Software and Data Technologies
122