![](bg2.png)
2009; Desailly et al., 2009). Force plates capture
the ground reaction force caused by the pressure ex-
erted by the foot on them. Thus by applying simple
thresholds on the recorded force, individual events of
HS and TO can be precisely identified (Mills et al.,
2007). Camera-based mocap systems track markers
that are placed on the body (Hanlon and Anderson,
2009; Aung et al., 2013). Generally, a laboratory
setup consists of a pair of force plates (one for each
foot) that is used concurrently with the mocap sys-
tem. Though these devices provide relatively pre-
cise measurements, they are highly expensive, immo-
bile and require competence in maintenance, opera-
tion and execution. The fact that they record the infor-
mation of only a couple of steps renders them inade-
quate for applications that need long-term or continu-
ous gait monitoring. For example, some authors have
hypothesized that in-coordination between the neu-
ral and locomotor control systems results in specific
gait patterns which can be used for analyzing neuro-
physiological disorder patients (Stolze et al., 2004;
Hausdorff, 2007) or older adults that have a fear of
falling (Herman et al., 2005; Reelick et al., 2009).
These applications would essentially require mobile
sensors and robust gait analysis algorithms for day-
to-day monitoring of such patients.
Pressure sensitive mats and foot switches provide
the option of being used outside the laboratory for
measuring gait events. The pressure sensitive mat
(e.g.GaitRite) consists of a grid of pressure sensors
that can be few meters in length (McDonough et al.,
2001). Foot switches like force sensitive resistors can
be attached at various positions below the feet or in-
soles of the shoe (Williamson and Andrews, 2000;
Aminian et al., 2002; Lau and Tong, 2008). How-
ever there have been studies to show that they are less
reliable and not durable over longer periods (Mans-
field and Lyons, 2003) and cannot differentiate be-
tween foot load changes due to walking and those
caused by weight shifting for non-walking tasks (Pap-
pas et al., 2001). Hence, from a long-term perspec-
tive, they might not be convenient for outdoor use in
daily life and be prone to mechanical failures. More-
over, these sensors provide only temporal information
and thus restrict the scope for any further analysis in-
volving spatial parameters of gait.
Consequently, for such long-term gait monitor-
ing and daily life applications, an alternative is to
use wearable inertial sensors like accelerometers and
gyroscopes. Advancements in MEMS technology
have made them miniature, low-powered, durable, in-
expensive, highly mobile and readily available (Ka-
vanagh and Menz, 2008). In recent years, many
gait event identification algorithms have been devel-
oped using inertial sensors (Rueterbories et al., 2010).
While some have used gyroscopes (Pappas et al.,
2001; Aminian et al., 2002; Lee and Park, 2011),
others have found it appropriate to use accelerome-
ters (Williamson and Andrews, 2000; Mansfield and
Lyons, 2003; Selles et al., 2005; Hanlon and Ander-
son, 2009; Sant’Anna and Wickstr¨om, 2010; Aung
et al., 2013). A major drawback of using inertial
sensors is that they provide highly noisy information
and thus require very robust algorithms for gait anal-
ysis. Based on the quality and type of information
they provide, some authors have discussed the pros
and cons of using either accelerometers or gyroscopes
(Aminian et al., 2002; Lau and Tong, 2008) though it
must be noted that these sensors are used and well re-
searched in other fields like aerospace and humanoids
(Khandelwal and Chevallereau, 2013). In the con-
text of gait event analysis, accelerometers seem to
be a better choice than gyroscopes for developing
automated gait event identification systems. Sudden
movements like jerks or turns during walking would
cause large gyro drift errors. Furthermore, gyroscopes
have high power consumption, long reaction time and
are very sensitive to temperature effects limiting their
long-term outdoor use (Tan and Park, 2005; Wood-
man, 2007). Accelerometers, on the other hand, suf-
fer from noise due to mechanical vibrations and cali-
bration errors but these do not diverge in time and in
many cases can be handled effectively.
Recently, many methods have been developed
for identifying gait events only from accelerometer
data. Some of these algorithms apply different tech-
niques to analyze signals obtained from individual
accelerometer axis (Williamson and Andrews, 2000;
Mansfield and Lyons, 2003; Selles et al., 2005; Tor-
realba et al., 2010). Hence, at the beginning of the
experiment, the accelerometer is positioned in some
specific orientation such that each accelerometer axis
is aligned with some pre-defined limb axis and the
assumption is made that the accelerometer shall stay
statically positioned during the entire movement of
the experiment. Otherwise, either the axis alignment
should be readjusted frequently or the exact orien-
tation of the accelerometer must be known through-
out the experiment (which might be difficult to es-
timate) to compensate for the misalignment of the
axes. A possible alternative could be to analyze the
resultant accelerometer signal instead as it is invari-
ant to individual axis alignment. Some other meth-
ods use machine learning techniques (Williamson and
Andrews, 2000; Aung et al., 2013) but the difficulty
with such algorithms is that they depend on labeled
training data and the addition or exclusion of any
parameter would require re-training the entire algo-
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
198