Signal Processing - March 2016 - 33

and, after that, a hierarchical HMM is applied to model the
correlation and constraints between commands. A continuous
HMM is proposed to jointly classify the pedestrian activity
and gait phases with the assumption that state-conditional
output density functions of the HMM to be a Gaussian mixture model [27]. This approach is robust to subject variability.
It will still perform well when new subject data is tested without any training for this subject. In-home assembly task recognition is performed using a HMM on accelerometer data
with fusion of the linear discriminant analysis (LDA) decision
from sound data [4]. A method for spotting sporadically
occurring gestures (e.g., handshake, drink, pick up the phone,
etc.) in a continuous data stream from body-worn inertial sensors was designed using a HMM [17]. The method contains
two stages. In the first stage, signal sections likely to contain
specific motion events are selected using a similarity searching algorithm; and in the second stage, the HMM is applied to
classify the activities.

certain optimization criterion. An adaptive filter is designed to
fuse all of the sensor information and pseudo-measurements
to provide a self-contained pedestrian tracking system during
normal walking [9]. In the cases that the systems are nonlinear
and the noise is non-Gaussian, a particle filter, which is more
complex, will usually perform better than a Kalman filter. A
particle filter is used to fuse the step length and heading information from inertial sensors to provide an indoor localization
system [12].

Performance analysis
The performance and efficiency of assisted living technologies can be evaluated using many metrics. The goal for this
section is to compare recent signal processing advances with
respect to accuracy, power consumption, and computational
complexity of the sensors and algorithms.

Accuracy

The accuracy of IMU-based signal processing techniques is a
key aspect for assisted living applications. The cost of faults
can be significant, especially when the techniques are used to
In addition to classification algorithms, advanced measureassist the elderly, individuals who are vulnerable, and those
ment models are applied to fuse different modalities of IMU
that are in need of care.
sensors (e.g., accelerometer, gyroscope, and magnetometer) to
Signal processing techniques are procompensate for errors and drifts. This leads
posed to reliably detect the human postural
to robust measurements for different tasks
Classification can be used transition and ADLs, recognize gestures,
in assisted living. Kalman filters and partito detect falls and prefalls, and track the users' sleeping patterns and
cle filters are among the most popular
diet. FFT was used to extract information
fusion techniques. The Kalman filter is an
to distinguish between
from IMU sensor data to recognize and disalgorithm that uses a model and a series of
healthy and unhealthy
tinguish falling, sitting, and walking activinoisy and possibly inaccurate measuremotor function, and to
ties [2]. Using FFT on data from a
ments observed over time to produce estidetect ADLs.
wrist-worn sensor with a 10-Hz sampling
mates of unknown variables that tend to be
rate was unable to accurately discern
more precise than those based on a single
between falling and sitting down. A method of physical activity
measurement alone. It is widely used in the navigation and
monitoring to detect activities such as sitting, standing, and
control systems. A conventional Kalman filter is used to
lying has sensitivities and specificities of 90.2% and 93.4% for
reduce the drift from inertial sensors in an indoor navigation
sitting, 92.2% and 92.1% for standing+walking, and, 98.4% and
system with foot-mounted strap-down inertial sensors [11].
99.7% for lying with a sternum-mounted sensor sampling at
The inertial navigation system calculates the position change
40 Hz [22]. Overall, the detection errors were 3.9% for standing +
at a high frequency rate, and the integration error from the
walking, 4.1% for sitting, and 0.3% for lying. Finally, the overinertial sensor will accumulate over time. The GPS is also a
all symmetric mean average errors were 12% for standing +
part of the system and when GPS data is available, the GPS
walking 8.2% for sitting, and 1.3% for lying. A model to fuse
derived positions are compared with the positions derived
data from hand movements and audio sampled at 2 kHz from a
from the inertial navigation system. The differences are fed
wood workshop to recognize workers' activities is presented
into a Kalman filter that estimates the errors from the inertial
[4]. Different methods were used to improve the classification
navigation system and compensates the measurements so that
and it is shown that in isolation, the accuracy of activity detecthe errors remain small. A Kalman filter is used to combine
tion is 98%, 87%, and 95% for the user-dependent, user-indethe acceleration, angular velocity and biomechanical conpendent, and user-adapted detection, respectively. A data set
straints to generate robust estimation of the knee joint flexion/
was created from a wrist-worn IMU sensor, and a method to
extension angles [20]. The gyroscope noise and the accelerdetect sleep and wake states was proposed [5]. The algorithm
ometer noise are modeled by the Kalman filter. The proposed
was compared with traditional algorithms using total sleep time
system works effectively for both walking and running for
(TST) and sleep efficiency (SE) as the comparison parameter.
five minutes when compared to a camera-based motion trackThe proposed method achieves an overall median accuracy of
ing system.
79% for detecting sleep and wake intervals.
Unlike the Kalman filter, the adaptive filter is a system
Several accurate human localization techniques are prowith a linear filter that has a transfer function controlled by
posed, leveraging IMU-based wearable solutions. An adaptive
variable parameters that are adaptively updated according to

Measurement models

IEEE SIgnal ProcESSIng MagazInE

March 2016

Table of Contents for the Digital Edition of Signal Processing - March 2016